Started in January,1974(Monthly)
Supervised and Sponsored by Chongqing Southwest Information Co., Ltd.
ISSN 1002-137X
CN 50-1075/TP
Current Issue
Volume 48 Issue 11A, 16 November 2021
Intelligent Computing
Overview of Nested Named Entity Recognition
YU Shi-yuan, GUO Shu-ming, HUANG Rui-yang, ZHANG Jian-peng, SU Ke
Computer Science. 2021, 48 (11A): 1-10.  doi:10.11896/jsjkx.201100165
Abstract PDF(2516KB) ( 2063 )   
References | Related Articles | Metrics
There are rich semantic relations and structural information between nested named entities,which is very important for the implementation of downstream tasks such as relation extraction and event extraction.The accuracy of text information extraction model has gradually exceeded the traditional rule-based method.Therefore,many scholars have carried out research on nested named entity recognition technology based on deep learning,and obtained the most advanced performance.This paper reviews the existing nested named entity recognition technology,This paper gives a comprehensive review of the existing nested named entity recognition technology,introduces the most representative methods and the latest application technology of nested named entity recognition,and discusses and prospects the challenges and development direction in the future.
Survey of Graph Neural Network in Community Detection
NING Yi-xin, XIE Hui, JIANG Huo-wen
Computer Science. 2021, 48 (11A): 11-16.  doi:10.11896/jsjkx.210500151
Abstract PDF(2363KB) ( 2540 )   
References | Related Articles | Metrics
Community structure is one of the universal topological properties in complex networks,and discovering community structure is the basic task of complex network analysis.The purpose of community detection is to divide the network into several substructures,which plays an important role in understanding the network and revealing its potential functions.Graph Neural Network (GNN) is a model for processing graph structure data,which has the advantage of feature extraction and representation from graph,and has become an important research field of artificial intelligence and big data.Network data is a typical graph structure data.Using graph neural network model to solve the problem of community detection is a new direction of community detection research.In this paper,we first discuss the GNN model,analyze the process of GNN community detection,and discuss the progress of existing GNN community detection and the direction of future research in detail from two aspects of overlapping community and non-overlapping community.
Chinese Commentary Text Research Status and Trend Analysis Based on CiteSpace
LI Jian-lan, PAN Yue, LI Xiao-cong, LIU Zi-wei, WANG Tian-yu
Computer Science. 2021, 48 (11A): 17-21.  doi:10.11896/jsjkx.210300172
Abstract PDF(2746KB) ( 1100 )   
References | Related Articles | Metrics
Natural language processing (NLP) has been a hot topic in the field of artificial intelligence (AI) recently,among which commentary-based text analysis has also attracts the attention of scholars.In this study,the research status and frontier development trend can be grasped through a visual analysis of the domestic literature on comment text analysis.A total of 453 valid core journal papers on the field are selected from CNKI as the data source.CiteSpace software is used to draw the knowledge map and analyze it.Analysis results show that the number of literature in this field has been on the rise in past 15 years.The cooperation among authors and among research institutions is not close,and a cohesive research group has not been formed.Sentiment analysis,online comments and deep learning are the main research hotspots at present.From the initial development of theoretical basis and the expansion of application direction,to the improvement of analysis methods and models in the later stage,scholars have gradually deepened the research in this field.In the future,the cooperative relationship between researchers and research institutions needs to be strengthened,and various models which are represented by deep learning will continue to develop and improve in the future.
Theoretical Research and Efficient Algorithm of Container Terminal Quay Crane Optimal Scheduling
GAO Xi, SUN Wei-wei
Computer Science. 2021, 48 (11A): 22-29.  doi:10.11896/jsjkx.201200167
Abstract PDF(2558KB) ( 1107 )   
References | Related Articles | Metrics
Quay crane scheduling problem is one of the most important scheduling problems in container terminals.The existing research results can not calculate the optimal scheduling of large-scale business in the feasible time,so the existing quay crane scheduling algorithms generally adopt heuristic strategies to ensure that scheduling can be calculated in the feasible time.In this paper,firstly,the correctness of the lower bound of completion time is proved theoretically,an optimal scheduling construction method is designed and the theoretical system of quay crane scheduling problem is completed.Secondly,based on the theoretical work,an algorithm of linear time complexity is designed to find the optimal scheduling.Finally,experiments show that the proposed method is significantly better than the existing methods in terms of solution quality and efficiency.
Multi-worker and Multi-task Path Planning Based on Improved Lion Evolutionary Algorithm forSpatial Crowdsourcing Platform
ZHAO Yang, NI Zhi-wei, ZHU Xu-hui, LIU Hao, RAN Jia-min
Computer Science. 2021, 48 (11A): 30-38.  doi:10.11896/jsjkx.201200085
Abstract PDF(3214KB) ( 853 )   
References | Related Articles | Metrics
In order to solve the problem of multi-worker and multi-task path planning for spatial crowdsourcing platform and aiming at solving the global optimal path planning scheme with the minimum time cost and distance cost,a path planning method based on the improved lion evolutionary algorithm is proposed.Firstly,a path planning model with task start and end points is proposed based on realistic problem scenarios.Secondly,by referring to the algorithm idea of lion evolutionary algorithm,the intelligent behavior of lions is improved,the expulsion behavior is introduced,and the chromosomal coding mode,crossover,mutation operation,etc.are designed for solving the problem.An improved lion evolutionary algorithm for multi-worker and multi-task path planning based on the spatial crowdsourcing platform is proposed.Finally,the improved lion evolutionary algorithm is used to solve the multi-worker and multi-task path planning model of the spatial crowdsourcing platform,and the problem is tested by making an example based on the real data set.The experimental results show the availability and effectiveness of the algorithm.
Bitcoin Price Forecast Based on Mixed LSTM Model
ZHANG Ning, FANG Jing-wen, ZHAO Yu-xuan
Computer Science. 2021, 48 (11A): 39-45.  doi:10.11896/jsjkx.210600124
Abstract PDF(4081KB) ( 1908 )   
References | Related Articles | Metrics
For the reason that Bitcoin price is highly nonlinear and non-stationary,this paper proposes four mixed forecasting model based on Long Short-Term Memory (LSTM) model to get better prediction performance.Firstly,we use Wavelet Transform (WT) and Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) to decompose and reconstruct the original sequence.Then,we introduce Sample Entropy (SE) to optimize the reconstruction.Finally,we predict the reconstructed sub-sequences respectively using LSTM and superpose the outcomes yields of prediction results.To evaluate the prediction performance,three evaluation functions are used,which are RMSE,MAPE and TIC.Besides,the results are compared with single LSTM model prediction result.The research shows that the prediction accuracy of mixed model is better than that of the single model,and the introduction of Sample Entropy can effectively reduce prediction error.
Traffic Prediction Model Based on Dual Path Information Spatial-Temporal Graph Convolutional Network
KANG Yan, XIE Si-yu, WANG Fei, KOU Yong-qi, XU Yu-long, WU Zhi-wei, LI Hao
Computer Science. 2021, 48 (11A): 46-51.  doi:10.11896/jsjkx.201200184
Abstract PDF(2472KB) ( 914 )   
References | Related Articles | Metrics
With the development of deep learning,neural network has a large number of applications in various fields,and intelligent transportation system is no exception.Traffic flow forecast is the cornerstone of intelligent traffic system and the core of the whole traffic forecast.In recent years,the use of the graph convolutional neural network has effectively improved the performance of traffic prediction.How to further improve the ability to capture the spatial and temporal characteristics of the graph will become a hot topic.In order to improve the accuracy of traffic prediction,this paper proposes a traffic prediction model based on the convolution network of dual path information spatial-temporal map.First of all,the traffic prediction model based on the graph convolution network has some shortcomings in long-distance dependence modeling,and has not fully mined the hidden relationship between the spatial-temporal diagram information and the missing information in the spatial-temporal diagram structure,so we propose a triple pooling attention mechanism to model the global context information.Based on the figure of each increase in parallel convolution layer and the time convolution triple pooling attention path,we construct a dual path information spatial-temporal convolution layer,enhance the generalization ability of convolution layer,improve the model's ability to capture long distance dependence,and spatial-temporal convolution layer can capture figure characteristics of space and time structure of spacetime,effectively improve the traffic prediction performance.Experimental results on two public transport data sets (METR-LA and PEMS-BAY) show that the proposed model has good performance.
Event Argument Extraction Using Gated Graph Convolution and Dynamic Dependency Pooling
WANG Shi-hao, WANG Zhong-qing, LI Shou-shan, ZHOU Guo-dong
Computer Science. 2021, 48 (11A): 52-56.  doi:10.11896/jsjkx.201200259
Abstract PDF(1770KB) ( 737 )   
References | Related Articles | Metrics
Event argument extraction is a very challenging subtask of event extraction.This task aims to extract the arguments in the event and the role they played.It is found that the semantic features and dependency features of sentences play a very important role in event argument extraction,and the existing methods often don't consider how to integrate them effectively.Therefore,this paper proposes an event argument extraction model using gated graph convolution and dynamic dependency pooling.This method uses BERT to extract the semantic features of sentences,and then two same graph convolution networks are used to extract the dependency features of sentences based on the dependency tree.The output of one graph convolution is used as the gating unit through the activation function.Then semantic features and dependency features are added and fused through gate unit.In addition,a dynamic dependency pooling layer is designed to pool the fused features.The experiment results on ACE2005 dataset show that the proposed model can effectively improve the performance of event argument extraction.
Fuzzy Reasoning Method Based on Axiomatic Fuzzy Sets
KANG Bo, PAN Xiao-dong, WANG Hu
Computer Science. 2021, 48 (11A): 57-62.  doi:10.11896/jsjkx.201200140
Abstract PDF(1702KB) ( 643 )   
References | Related Articles | Metrics
Based on the axiomatic fuzzy set theory,this paper regards fuzzy reasoning as the mapping between two fuzzy membership spaces,and gives three basic forms of fuzzy reasoning output results by using the composition of input fuzzy sets in fuzzy membership spaces.For strongly negative operators,t-modulus operators and t-comodule operators,the perturbation of these operators in fuzzy membership space is discussed by using Minkowski integral distance,and the continuity of fuzzy reasoning method proposed in this paper is analyzed on this basis.
New Optimization Mechanism:Rain
LIU Hua-ling, PI Chang-peng, LIU Meng-yao, TANG Xin
Computer Science. 2021, 48 (11A): 63-70.  doi:10.11896/jsjkx.201100032
Abstract PDF(3797KB) ( 737 )   
References | Related Articles | Metrics
The loss function of the traditional model in the field of machine learning is convex,so it has a global optimal solution.The optimal solution can be obtained through the traditional gradient descent algorithm (SGD).However,in the field of deep learning,due to the implicit expression of the model function and the interchangeability of neurons in the same layer,the loss function is a non-convex function.Traditional gradient descent algorithms cannot find the optimal solution,even the more advanced optimization algorithms such as SGDM,Adam,Adagrad,and RMSprop cannot escape the limitations of local optimal solutions.Although the convergence speed has been greatly improved,they still cannot meet the actual needs.A series of existing optimization algorithms are improved based on the defects or limitations of the previous optimization algorithms,and the optimization effect is slightly improved,but the performance of different data sets is inconsistent.This article proposes a new optimization mechanism Rain,which combines the dropout mechanism in deep neural networks and integrates it into the optimization algorithm to achieve.This mechanism is not an improved version of the original optimization algorithm.It is a third-party mechanism independent of all optimization algorithms,but it can be used in combination with all optimization algorithms to improve its adaptability to data sets.This mechanism aims to optimize the performance of the model on the training set.The generalization problem on the test set is not the focus of this mechanism.This article uses Deep Crossing and FM two models with five optimization algorithms to conduct experiments on the Frappe and MovieLens data sets respectively.The results show that the model with the Rain mechanism has a significant reduction in the loss function value on the training set,and the convergence speed is accelerated,but its performance on the test set is almost the same as the original model,that is,its generalization is poor.
Risk Control Model and Algorithm Based on AP-Entropy Selection Ensemble
WANG Mao-guang, YANG Hang
Computer Science. 2021, 48 (11A): 71-76.  doi:10.11896/jsjkx.210200110
Abstract PDF(2094KB) ( 770 )   
References | Related Articles | Metrics
In recent years,many risk control problems have emerged in the field of Internet finance.For this,we adopt a variety of feature selection methods to preprocess data indicators in the field of risk control,and construct a comprehensive risk control indicator system for corporate credit.And we use stacking ensemble strategy to study credit risk model based on AP-entropy.There are two layers of learners in credit risk model.The idea of selection ensemble is introduced to select the base learners from the category and quantity.First,in machine learning algorithms such as Logistic regression,back propagation neural network,AdaBoost,AP clustering algorithm is used to select a heterogeneous learner suitable for corporate credit risk as the base learner.Se-condly,in each iteration of the learner,entropy is used to select the best learner,and the base learner with the highest F1 value is automatically selected.Among them,the improved algorithm based on entropy improves the efficiency of base learner selection process and reduces the computational cost of the model.Xgboost is selected as the secondary base learner.The empirical results show that the proposed model has good performance and generalization ability.
Route Planning of Unstructured Road Including Repeat Node Based on Bidirectional Search
CAO Bo, CHEN Feng, CHENG Jing, LI Hua, LI Yong-le
Computer Science. 2021, 48 (11A): 77-80.  doi:10.11896/jsjkx.201200193
Abstract PDF(2415KB) ( 838 )   
References | Related Articles | Metrics
Aiming at high-precision navigation route planning around unstructured environment with repeat node,firstly,we propose a map model construction method based on all-direction intersection structure,taking the turn constraint into account.Based on the traditional navigation map,the intersection structure is refined more detailed with the navigation node set based on turn,and the turning constraint is processed to ensure that the turning constraint are met when topology relation is formed between different nodes.Then,A* algorithm based on bidirectional search is designed according to the model to solve the route planning problem of unstructured road with repeated node,which expands the route search from the origin and the destination meanwhile until the optimal route is obtained.Finally,a comparative experiment is carried out around a field environment.Results show that route obtained based on the map model in this paper can satisfy the restriction of turning constraint and effectively solve the problem of repeat node route planning.
Deep Clustering Model Based on Fusion Variational Graph Attention Self-encoder
KANG Yan, KOU Yong-qi, XIE Si-yu, WANG Fei, ZHANG Lan, WU Zhi-wei, LI Hao
Computer Science. 2021, 48 (11A): 81-87.  doi:10.11896/jsjkx.210300036
Abstract PDF(3405KB) ( 976 )   
References | Related Articles | Metrics
As one of the most basic tasks in data mining and machine learning,clustering is widely used in various real-world tasks.With the development of deep learning deep clustering has become a research hotspot.Existing deep clustering algorithms are mainly from two aspects of node representation learning or structural representation learning.Less work considers fusing these two kinds of information at the same time to complete representation learning.This paper proposes a deep clustering model FVGTAEDC (Deep Clustering Model Based on Fusion Varitional Graph Attention Self-encoder),this model joints the autoencoderand the variational graph attention autoencoder for clustering.In the model,the autoencoder integrates the variational graph attention autoencoder from the network to learn (low-order and high-order) structural representations,and then the feature representation is learned from the original data.While the two modules are trained,in order to adapt to the clustering task,self-supervised clustering training for the autoencoder module is integrated with the representation features of the node and the structure information.Comprehensive clustering loss,autoencoder reconstruction data loss,and variational graph attention autoencoder reconstruction adjacency matrix loss,the relative entropy loss of the posterior probability distribution and the prior probability distribution.The method can effectively aggregate the attributes of nodes and the structure of the network,while optimizing the assignment of cluster labels and learning the representation features suitable for clustering.Comprehensive experiments prove that the method is better than the current advanced deep clustering method on 5 real data.
Multiple Fault Localization Method Based on Deep Convolutional Network
Computer Science. 2021, 48 (11A): 88-92.  doi:10.11896/jsjkx.210200096
Abstract PDF(2481KB) ( 645 )   
References | Related Articles | Metrics
Most of the current fault localization methods solve single fault localization,but the faults are related to each other.How to find the relationship between these faults and the test results and the relationship between the faults,and reduce the impact of coincidental correct test cases and similar test cases on the suspiciousness of sentences is very important to improve the efficiency of multiple fault localization.In order to solve the above problems,this paper proposes a multiple fault localization method based on deep convolutional network.A set of suspiciousness with high accuracy is obtained through a deep convolutional network with a special structure,and then applied to forward slicing and backward slicing,the correlation between faults and faults is found to locate multiple faults.Experiments show that the multiple fault localization efficiency of the method in this paper is stronger than that of the existing classic fault localization methods.
Hybrid Artificial Chemical Reaction Optimization with Wolf Colony Algorithm for Feature Selection
ZHANG Ya-chuan, LI Hao, SONG Chen-ming, BU Rong-jing, WANG Hai-ning, KANG Yan
Computer Science. 2021, 48 (11A): 93-101.  doi:10.11896/jsjkx.210100067
Abstract PDF(2717KB) ( 596 )   
References | Related Articles | Metrics
Wrapper feature selection is a data preprocessing method for reducing original dataset dimensionality by screening the most informative features to maximize the classification accuracy synchronously.In order to improve the wrapper feature selection ability,this paper proposes a hybrid artificial chemical reaction wolf colony optimization algorithm for selecting feature-ACR-WCA.First,ACR-WCA algorithm adopts natural strategy,imitates the search strategy of wolves,so can quickly approach the solution space.Secondly,in order to deal with data features effectively,the S-shaped transfer function is used to deal with binary features in the initialization stage.Then the fitness function of the algorithm is given by combining classification accuracy and the number of features.Meanwhile,the method uses K-Nearest Neighbor (KNN) classifier for training and tested data by K-fold cross-validation to overcome the over fitting problem.The experiments are trained based on 21 famous and different dimensionality dataset,and compared with four traditional methods and three nearly methods.Experimental results show that the algorithm is efficient and reliable.It can select the most features for classifications tasks with high accuracy.
Incremental Tag Propagation Algorithm Based on Three-way Decision
XIN Xian-wei, SHI Chun-lei, HAN Yu-qi, XUE Zhan-ao, SONG Ji-hua
Computer Science. 2021, 48 (11A): 102-105.  doi:10.11896/jsjkx.210300065
Abstract PDF(2410KB) ( 552 )   
References | Related Articles | Metrics
As a new method of granular computing,the three-way decision(3WD) has unique advantages in dealing with uncertain and imprecise problems.Aiming at the high random uncertainty and redundancy of the label propagation algorithm (LPA) in the node update process,an incremental label propagation algorithm based on the three-way decision (3WD_ILPA) is proposed.First,the concept and calculation method of adjacency fuzzy information measure are given and used to generate the probability transfer matrix between any two nodes.Then,the three-way decision is integrated into the dynamic update process,and the node with the highest precision is added to the next periodic iteration until convergence.Furthermore,the algorithm flow of 3WD_ILPA is given in detail.Finally,the autism (ASD) recognition experiment is carried out on the ABIDE data set.By comparing with traditional machine learning,deep learning and transfer learning methods,the results show that the proposed method has higher accuracy.
Optimization of Sharing Bicycle Density Distribution Based on Improved Salp Swarm Algorithm
ZHOU Chuan
Computer Science. 2021, 48 (11A): 106-110.  doi:10.11896/jsjkx.210700096
Abstract PDF(2319KB) ( 492 )   
References | Related Articles | Metrics
In this article,an improved sea-squirt algorithm is proposed for the urban bike-sharing distribution density optimization problem.First,the sharing bicycle distribution density optimization problem is converted into a functional optimization problem,and the objective function of optimization is established with waiting time,time spent,cost and safety cost as evaluation indexes.Secondly,a one-dimensional normal cloud model and a nonlinear decreasing control strategy are introduced to improve the leader search mechanism in the Bottleneck algorithm to enhance the mining ability of local data;an adaptive strategy is introduced to improve the follower search mechanism of the original algorithm to avoid the algorithm falling into the local optimum.Finally,the effectiveness of the proposed optimization algorithm is verified by the standard test function and the simulation of shared bicycle distribution density.The results show that the improved Bottlenose sheath algorithm has better stability and global search capability than the original algorithm,firefly algorithm and artificial bee colony algorithm,and can better optimize the distribution density of shared bicycles and improve the regional utilization rate of shared bicycles,which is a reference value for the development of intelligent transportation.It has certain reference value for the development of intelligent transportation.
Heterogeneous Network Link Prediction Model Based on Supervised Learning
HUANG Shou-meng
Computer Science. 2021, 48 (11A): 111-116.  doi:10.11896/jsjkx.210300030
Abstract PDF(2287KB) ( 618 )   
References | Related Articles | Metrics
The research on traditional heterogeneous network link prediction has path-predicted algorithm and MPBP(meta-path feature-based backpPropagation neural network model) algorithm based on the metapath supervised learning.However,they can't make full use of the rich information provided by heterogeneous network to make link prediction.Based on the traditional supervised learning algorithm,this paper first designs the HLE-T(heterogeneous link entropy with time) algorithm in order to increase the link entropy and time dynamic information.Moreover,it constructs the MSLP(modified supervised link prediction)model of the Supervised learning algorithm with the multi-classification problem by the numerical segment of the link strength and weak relationship,and finally completes the experimental test under four data sets with different density.The experimental results show that the MSLP model improves the link prediction performance in heterogeneous network to some extent,and has some reference significance for the future link prediction research.
Space Retrieval Method to Retrieve Straight Line for Vector Line Data
LIU Ze-bang, CHEN Luo, YANG An-ran, LI Si-jie
Computer Science. 2021, 48 (11A): 117-123.  doi:10.11896/jsjkx.210100084
Abstract PDF(3786KB) ( 543 )   
References | Related Articles | Metrics
Shape cognition is one of the basic problems of spatial cognition.As the most basic shape- line,the retrieval based on it has important research significance in equipment layout,route planning and vehicle testing.Aiming at the problem of low efficiency and low accuracy in line recognition of remote sensing image by traditional methods,this paper presents a space retrieval method to retrieve straight line for vector line data.Firstly,in order to describe the flatness of line elements,the concepts of “flatness information” are defined.Then the subsection model of the flat sequence of line elements is established,the line elements are decomposed into a set of relatively straight subsegment sequences.Combined with the above two parts,the flatness information of subsegment is calculated after segmenting the line elements,and the final retrieval results are obtained by combining the retrieval conditions.Taking OSM roads network data as the research object,comparative tests verify that the retrieval method is faster,the retrieval results are more fully,and straight line roads search results are consistent with people's cognitive in shape.Moreover,the proportion of high-grade roads is 71.1%,and the proportion of small roads is only 2.8%,which is also consistent with the cognition of the property of straight roads in reality,and verifies the feasibility and rationality of the method.
Rule-based Automatic Recognition of Relations for Marked Complex Sentences
YANG Jin-cai, HU Qiao-ling, HU Quan
Computer Science. 2021, 48 (11A): 124-129.  doi:10.11896/jsjkx.210100226
Abstract PDF(1864KB) ( 757 )   
References | Related Articles | Metrics
Semantic expression of Chinese complex sentences is complicated.As an important content of Chinese discourse studies,complex sentences classification has always been a hot spot in the field of natural language processing.This paper summarizes and excavates more than ten types of literal and syntactic features for automatic identification of complex sentence categories,formalizes the features and constitutes rules,and uses the mechanism of relational words to trigger the rules to identify twelve types of relationship categories for marked complex sentences.Experimental results show that this method has achieved a higher accuracy rate,which is better than the existing methods.
Joint Learning of Causality and Spatio-Temporal Graph Convolutional Network for Skeleton- based Action Recognition
YE Song-tao, ZHOU Yang-zheng, FAN Hong-jie, CHEN Zheng-lei
Computer Science. 2021, 48 (11A): 130-135.  doi:10.11896/jsjkx.201200205
Abstract PDF(2636KB) ( 847 )   
References | Related Articles | Metrics
In recent years,skeleton based human action recognition has attracted extensive attention due to its simplicity and robustness.Most of the skeleton based human action recognition methods,such as spatio-temporal graph convolutional network (ST-GCN),distinguish different actions by extracting the temporal features of consecutive frames and the spatial features of skele-ton joints within frames,achieve good results.In this paper,considering the causality of human action,we propose an action recog-nition method combining causality and spatio-temporal graph convolutional network.In view of the complexity of obtaining weight,we propose a method to calculate joint weight based on causality.According to the causality,we assign weights to skeleton graph,and use weights as auxiliary information to enhance graph convolutional network to improve the weight of some joints with strong driving force in the neural network,so as to reduce the attention of low importance joints and enhance the attention of high importance joints.Compared with ST-GCN,our methodimproves the recognition accuracy of both Top-1 and Top-5,and the recognition accuracy reaches 97.38% (Top-1) and 99.79% (Top-5) on the real TaiChi dataset,which strongly prove that our method can effectively learn and enhance the discriminative features.
End-to-End Chinese-Braille Automatic Conversion Based on Transformer
JIANG Qi, SU Wei, XIE Ying, ZHOUHONG An-ping, ZHANG Jiu-wen, CAI Chuan
Computer Science. 2021, 48 (11A): 136-141.  doi:10.11896/jsjkx.210100025
Abstract PDF(2481KB) ( 1025 )   
References | Related Articles | Metrics
Chinese-Braille automatic conversion concerns the life and learning of 17 million visually impaired people in China and the national information accessibility construction.All existing Chinese-Braille conversion methods adopt multi-step process,which firstly segment Chinese text according to Braille word segmentation rules,then mark tone for Chinese characters.This paper studies end-to-end deep learning system that directly converts Chinese into Braille.The encoder-decoder model transformer is trained on Chinese-Braille corpus.Based on six-month data of People's Daily,totaling about 12 million characters,this paper builds three Chinese-Braille corpora of Chinese common Braille,current Braille and Chinese double-phonic Braille systems.The experimental results show that the method proposed in this paper can convert Chinese into Braille in one step,and reaches BLEU score of 80.25%,79.08% and 79.29% in Chinese common Braille,current Braille and Chinese double-phonic Braille.Compared with the existing methods,this method requires a corpus which is less difficult to construct and the engineering complexity is lower.
Optimization of Empire Competition Algorithm Based on Gauss-Cauchy Mutation
WEI Xin, FENG Feng
Computer Science. 2021, 48 (11A): 142-146.  doi:10.11896/jsjkx.201200071
Abstract PDF(2199KB) ( 669 )   
References | Related Articles | Metrics
In order to solve the problems of slow convergence speed and easy to fall into local optimum in the competition process of imperial competitive algorithm (ICA),a new imperial competitive algorithm based on Gauss-Cauchy mutation (GCICA) is proposed.Gauss mutation is introduced in ICA Empire competition to speed up the convergence speed in the competition process;after the Empire perishes,the diversity is reduced and only in a small area for optimization,and Cauchy mutation is introduced to make it jump out of local optimum.By analyzing the simulation results of the algorithm with Gauss,Cauchy and Gauss-Cauchy mutation on several typical benchmark functions,the convergence speed and optimization accuracy of GCICA are improved.
Real-time Performance Analysis of Intelligent Unmanned Vehicle System Based on Absorbing Markov Chain
WU Pei-pei, WU Zhao-xian, TANG Wen-bing
Computer Science. 2021, 48 (11A): 147-153.  doi:10.11896/jsjkx.210300050
Abstract PDF(2266KB) ( 580 )   
References | Related Articles | Metrics
With the advancements of artificial intelligence technology and the development of human-cyber-physical systems,intelligent unmanned vehicle systems are becoming the forefront of the new generation of artificial intelligence research.The intelligent unmanned vehicle system performs real-time decision based on vehicle and environmental data to control the unmanned vehicle.Therefore,the intelligent unmanned vehicle system has high real-time performance requirements.Analysis of the real-time performance of the system is one of the methods to ensure the safety and reliability of this kind of system.In order to analyze the real-time performance of the intelligent unmanned vehicle system,this paper takes the intelligent unmanned vehicle lane changing system as a scenario.First,the MARTE model is used to model the intelligent unmanned vehicle lane changing system,and the performance requirements parameters are added in the early system design.Then,through model transformation,the MARTE model is transformed into an absorption Markov chain.Finally,the relevant theories and formulas of the absorption Markov chain are used to comprehensively estimate the real-time performance indicators of the intelligent unmanned vehicle system,and analyze the key modules that affect the real-time performance of the entire system.The experimental results show that the model and analysis method proposed in the article can better analyze the real-time performance of the intelligent unmanned vehicle system.The analysis found that the accuracy and response time of the intelligent modules in the system restrict each other,and it is necessary to find a balance between the two in different operating scenarios to obtain better real-time performance.
Chinese Ship Fault Relation Extraction Method Based on Bidirectional GRU Neural Network and Attention Mechanism
HOU Tong-jia, ZHOU Liang
Computer Science. 2021, 48 (11A): 154-158.  doi:10.11896/jsjkx.210100215
Abstract PDF(2121KB) ( 979 )   
References | Related Articles | Metrics
With the development of deep learning,more and more deep learning models are applied to relational extraction tasks.The traditional deep learning model can not solve the long-distance learning task,and the performance of the traditional deep learning model is worse when the noise of text extraction is large.To solve the above two problems,a deep learning model based on bidirectional GRU (gated recurrent unit) neural network and attention mechanism is proposed to extract the relationship between Chinese ship faults.Firstly,by using bidirectional GRU neural network to extract text features,the problem of long dependence of text is solved,and the running time loss and iteration times of the model are also reduced.Secondly,by establishing sentence level attention mechanism,the negative impact of noisy sentences on the whole relationship extraction is reduced.Finally,the model is trained on the training set,and the accuracy,recall,and F1 values are calculated on the real test set to compare the model with existing methods.
Dual Autoregressive Components Traffic Prediction Based on Improved Graph WaveNet
LI Hao, WANG Fei, XIE Si-yu, KOU Yong-qi, ZHANG Lan, YANG Bing, KANG Yan
Computer Science. 2021, 48 (11A): 159-165.  doi:10.11896/jsjkx.201200051
Abstract PDF(2482KB) ( 556 )   
References | Related Articles | Metrics
With the construction of smart cities,urban traffic flow forecasting is crucial in intelligent traffic early warning and traffic management decision-making.Due to the complex temporal and spatial correlation,it is a challenge to effectively predict traffic flow.Most of the existing traffic flow prediction methods use machine learning algorithms or deep learning models,and the methods have their own advantages and disadvantages.If the two advantages can be combined,the accuracy of traffic flow prediction will be further improved.Aiming at the traffic spatio-temporal data,a dual autoregressive component traffic prediction model based on improved Graph WaveNet is proposed.First,the three time convolution layers are effectively fused through the gated three-branch time convolution network,which further improves the ability of capture time correlation.Second,the autoregression component is introduced for the first time to effectively fuse the autoregression component with the gated three-branch time convolution network and the convolution layer,so that the model can fully reflect the linear and non-linear relationship between space-time data.Through experiments on two real public transportation data sets of METR-LA and PEMS-BAY,the proposed model is compared with other traffic flow prediction benchmark models.The results show that whether it is a short-term or long-term forecast,the model proposed in the paper is better than the benchmark model in all indicators.
Spatial Cylinder Fitting Based on Projection Roundness and Genetic Algorithm
GAO Shuai, XIA Liang-bin, SHENG Liang, DU Hong-liang, YUAN Yuan, HAN He-tong
Computer Science. 2021, 48 (11A): 166-169.  doi:10.11896/jsjkx.201100057
Abstract PDF(2266KB) ( 807 )   
References | Related Articles | Metrics
In order to solve the problem of strong nonlinearity,poor robustness and weak stability of spatial cylindrical surface,a method of fitting spatial cylindrical surface based on projection roundness and genetic algorithm is proposed.First of all,the projection roundness of a cylindrical surface on a plane is calculated by coordinate transformation.Then,the normal vectors of the plane in which the optimal projection roundness lies are searched quickly by means of the global optimization characteristic of genetic algorithm,and then the radius and axis equation of the cylindrical surface in space are calculated by plane projection.Finally,the inverse coordinate transformation is used to obtain the spatial cylindrical surface feature parameters under the original coordinates.Since the projection roundness of the cylindrical surface is only related to the direction angle of the normal vector of the projection plane,the solution space can cover all the spatial cylindrical surface of the pose only by reasonably setting the range of the direction angle of the normal vector.In addition,the genetic algorithm has good global optimization and convergence,and can achieve good results in spatial cylindrical surfaces and such multivariable nonlinear optimization problems.Simulation experiments and practical applications show that compared with the traditional fitting method,the spatial cylindrical fitting method based on projection roundness and genetic algorithm has the advantages of no need to estimate initial value,strong robustness,high fitting precision and good stability.This algorithm is an effective method to achieve cylinder fitting of arbitrary pose space.
Chinese Text Classification Model Based on Improved TF-IDF and ABLCNN
JING Li, HE Ting-ting
Computer Science. 2021, 48 (11A): 170-175.  doi:10.11896/jsjkx.210100232
Abstract PDF(2431KB) ( 693 )   
References | Related Articles | Metrics
Text classification which is often used in information retrieval,emotion analysis and other fields,is a very important content in the field of natural language processing and has become a research hotspot of many scholars.Traditional text classification model exists the problems of incomplete text feature extraction and weak semantic expression,thus,a text classification model based on improved TF-IDF algorithm and attention base on Bi-LSTM and CNN (ABLCNN) is proposed.Firstly,the TF-IDF algorithm is improved by using the distribution relationship of feature items within and between classes and location information to highlight the importance of feature items,the text is represented by word vector trained by word2vec tool and improved TF-IDF.Then,ABLCNN extracts the text features.ABLCNN combines the advantages of attention mechanism,long-term memory network and convolutional neural network.ABLCNN not only extracts major the context semantic features of the text,but also takes into account the local semantic features,At last,the feature vector is classified by softmax function.Chinese text classification model based on improved TF-IDF and ABLCNN is tested on THUCNews dataset and online_ shopping_ 10_cats dataset.The results of experimental show that the accuracy on the THUCNews dataset is 97.38% and the accuracy on the online_ shopping_ 10_cats dataset is 91.33%,the accuracy of experiment is higher than that of other text classification models.
Big Data & Data Science
Point-of-interest Recommendation:A Survey
XING Chang-zheng, ZHU Jin-xia, MENG Xiang-fu, QI Xue-yue, ZHU Yao, ZHANG Feng, YANG Yi-ming
Computer Science. 2021, 48 (11A): 176-183.  doi:10.11896/jsjkx.201100021
Abstract PDF(2522KB) ( 1623 )   
References | Related Articles | Metrics
Point-of-interest (POI) recommendation is an important service in location-based social networks (LBSN),which has an important impact on both merchants and customers.As a typical example of spatio-temporal data,POI recommendation has been widely concerned,so it has become a hot research topic in academic circles in recent years.This paper analyzes the influencing factors of interest point recommendation,summarizes the traditional methods of interest point recommendation,the latest graph-based embedding methods and the application of graph neural networks in the field of point-of-interest recommendation are analyzed.Finally,it analyzes the challenges faced by points of interest recommendation and future research trends.
Key Technologies and Development Trends of Big Data Trade Based on Blockchain
CAO Meng, YU Yang, LIANG Ying, SHI Hong-zhou
Computer Science. 2021, 48 (11A): 184-190.  doi:10.11896/jsjkx.210100163
Abstract PDF(1870KB) ( 1997 )   
References | Related Articles | Metrics
In the era of big data,the value of data has become increasingly prominent,and the needs of big data trading for diffe-rent entities are becoming more and more urgent.Traditional big data trading based on centralized platforms have many risks such as malicious collection of user data,privacy leakage,data resale,and data fraud.It is generally believed that using blockchain technology with the characteristics of decentralization,transparency,privacy protection,and tamper resistance is an important way to solve the above problems in big data trading.However,the development of blockchain technology in big data trading is still in the early stage,and the application plans are not yet mature.In this regard,this paper summarizes the current data trading models based on blockchain technology proposed by the academic community,introduces the methods of improving centralized big data trading from the three perspectives of privacy protection,data resale and trading fairness by using blockchain technology,and analyzes the advantages and disadvantages of each method.Then,this paper analyzes the current challenges and future development directions of “blockchain+big data trading” from the aspects of privacy protection,identity authentication,massive data,and va-lue transfer.
Trajectory Next Footprint Prediction Model Based on Adaptive Timestamp and Multi-scale Feature Extraction
LI Ai-ling, ZHANG Feng-li, GAO Qiang, WANG Rui-jin
Computer Science. 2021, 48 (11A): 191-197.  doi:10.11896/jsjkx.201200015
Abstract PDF(2442KB) ( 757 )   
References | Related Articles | Metrics
Location-based services have become a part of human life style,and various mobile terminal devices generate a large amount of temporal and spatial contextual user information,which can be used to predict the user's next footprint.Some solutions have been proposed to predict the user's next footprint,including recursive motion function (RMF),matrix factorization (MF),differential autoregressive moving average model (ARIMA),Markov chain (MC),and personalization Markov chain (FPMC),Kalman filter (KF),Gaussian mixture model and tensor decomposition (TF).In addition,deep neural network methods such as ST-RNN,POI2Vec,DeepMove,VANext,etc.can also be used to predict the user's next footprint.These methods use recurrent neural networks (RNN) to capture sequential motion patterns from human activities.However,existing methods use some artificially set thresholds to segment human mobility data for user movement pattern learning.The artificial fixed time stamp setting not only introduces human subjective factors,but also ignores the differences between different users.It may lead to deviations in the movement pattern.The existing methods for the extraction of user trajectory features are too singular,and a single feature ignores a lot of potential user trajectory information.The trajectory prediction model based on adaptive timestamp and multi-scale feature extraction (AMSNext) aims to combine the time statistical characteristics of historical trajectory data for the first time,adaptively define a personalized timestamp for each user,and focus on the differences between different user motion modes.Combined with time series feature extraction methods to extract user trajectory features at multiple scales,and at the same time,to achieve multi-scale feature dimension unification,normalized causal embedding will be used to embed features in vector.Experiments show that the model can achieve higher prediction accuracy.
Deep Community Detection Algorithm Based on Network Representation Learning
PAN Yu, ZOU Jun-hua, WANG Shuai-hui, HU Gu-yu, PAN Zhi-song
Computer Science. 2021, 48 (11A): 198-203.  doi:10.11896/jsjkx.210200113
Abstract PDF(2251KB) ( 737 )   
References | Related Articles | Metrics
Mining the community structure in the complex network is helpful to understand the internal structure and functional characteristics of the network,which has important theoretical value and significant practical significance.With the rapid development of information technology,the explosive growth of network data poses an unprecedented challenge for community detection.In this paper,the deep neural network is utilized to connect network representation learning and community detection domains,and a deep community detection method based on network representation learning is proposed.Firstly,the structural closeness of nodes is quantified according to their potential community membership similarities,and then a novel community structure method is proposed to construct the community structure matrix.Furthermore,a deep autoencoder that has several layers with non-linear functions is developed.The community structure matrix is used as the input of the deep autoencoder to obtain the low-dimension representation of the nodes which preserve the potential community structure.Finally,the K-means clustering strategy is applied to the network representation to obtain the community structure.Extensive experiments on both synthetic and real-world datasets of different scales demonstrate that the proposed method is feasible and effective.
Adaptive Frequency Domain Model for Multivariate Time Series Forecasting
WANG Xiao-di, LIU Xin, YU Xiao
Computer Science. 2021, 48 (11A): 204-210.  doi:10.11896/jsjkx.210500129
Abstract PDF(3528KB) ( 1420 )   
References | Related Articles | Metrics
In recent years,the research enthusiasm for time series data in academic and industrial fields has been increasing,but the frequency information contained in it still lacks effective modeling.The studies found that time series forecasting relies on different frequency patterns,providing useful clues for future trend forecasting:short-term series forecasting relies more on high-frequency components,while long-term forecasting focuses more on low-frequency data.In order to better mine the multi-frequency mode of time series,this paper proposes a multi-feature adaptive frequency domain prediction model (MAFD).MAFD is composed of two stages.In the first stage,it uses XGBoost algorithm to measure the importance of the input vector and selects high-importance features.In the second stage,the model integrates the frequency feature extraction of the time series and the frequency domain modeling of the target sequence,and builds an end-to-end prediction network based on the dependence of the time series on the frequency mode.The innovation of MAFD is reflected in the predictive network's ability to automatically focus on diffe-rent frequency components according to the dynamic evolution of the input sequence,thereby revealing the multi-frequency pattern of the time series and strengthening the learning ability of the model.This work uses 4 datasets from different fields to verify the performance of the model.The experimental results show that compared with the existing classic prediction models,MAFD has higher accuracy and smaller lag.
Link Prediction Algorithm Based on Ego Networks Structure and Network Representation Learning
ZHAO Man, ZHAO Jia-kun, LIU Jin-nuo
Computer Science. 2021, 48 (11A): 211-217.  doi:10.11896/jsjkx.201200231
Abstract PDF(2249KB) ( 717 )   
References | Related Articles | Metrics
Link prediction is a research direction that has attracted much attention in the field of network analysis and mining.The link prediction algorithm predicts the missing connections in the network,which is actually a process of data mining,and the inferred possible future connections are related to the development and evolution of the network.Therefore,how to improve the accuracy of link prediction is a meaningful and challenging research.Based on the latest research on ego network decomposition and community clustering,a link prediction algorithm (Ego-Embedding) is proposed,which is based on ego network structure characteristics and network representation learning.Ego-Embedding converts the original network into a persona graph,and then reconstructs the embedding process by combining the microstructure information and context information of the network,learning one or more vector representations for each node,so that the vector representation can describe the node information more accurately,thereby improving the accuracy of link prediction.This paper conducts experimental simulations on three public data sets (Facebook,PPI-Yeast,ca-HepTh),and uses AUC as the evaluation index.The experimental results show that the performance of the algorithm Ego-Embedding is better than the five experimental comparison methods (CN,AA,node2vec,M-NMF,Splitter),and the best link prediction AUC reduces the error by up to 47%.
Web Page Wrapper Adaptation Based on Feature Similarity Calculation
CHEN Ying-ren, GUO Ying-nan, GUO Xiang, NI Yi-tao, CHEN Xing
Computer Science. 2021, 48 (11A): 218-224.  doi:10.11896/jsjkx.210100230
Abstract PDF(2050KB) ( 568 )   
References | Related Articles | Metrics
With the development of big data,Internet data has exploded.As an important information carrier,the Web contains various types of information.The wrapper is proposed to extract target data from messy Web information.However,with frequent Web page updates,minor structural changes may cause the original wrapper to fail,leading to increased maintenance costs for the wrapper.Aiming at the robustness and maintenance cost of the wrapper,a Web page wrapper adaptive technology based on feature similarity calculation is proposed.This technology mainly analyzes the feature set of the new Web page and the feature information contained in the old wrapper,and calculates the similarity of the Web page to relocate the mapping area and mapping data items of the old wrapper in the new Web page,and make the old wrapper based on the mapping relationship able to adapt the data extraction of new Web pages.The technology is mainly used for experiments on various types of Websites,including shopping,news,information,forums and services.250 pairs of old and new versions of Web pages,totaling 500 Web pages,are selected for wrapper adaptation experiments.The experimental results show that when the Web page structure changes,the method can effectively adapt to the data extraction of the new Web page,and the average precision and average recall of data extraction reach 82.2% and 84.36%,respectively.
Analysis of Workload Failure in Co-located Data Centers
JIANG Cong-feng, YIN Ji-liang, HU Hai-zhou, YAN Long-chuan, ZHANG Ji-lin, WAN Jian, QIU Ye-liang
Computer Science. 2021, 48 (11A): 225-231.  doi:10.11896/jsjkx.201200066
Abstract PDF(3549KB) ( 735 )   
References | Related Articles | Metrics
Datacenter workload co-location can greatly increase the resource utilization of cloud data centers,while it also increases the scheduling complexity and job failures.In this paper,the cluster trace dataset from Alibaba Cloud is investigated,and the characteristics of batch workload failure rates and cluster resource utilization are studied.The main contributions and findings of this paper are as follows.First,Only a small portion of specific types of jobs account for the overall cluster failure rate and resource waste due to job failures.Second,the execution time of task failover in the Fuxi distributed scheduler can be quantified as Gaussian distribution,and the task scheduling delay can be quantified as Zipf distribution.Third,Based on the failed instances distribution on cluster nodes,it's found that the job failures randomly occur in the cluster spatially,and the failed jobs are prone to fail again after their failovers.Moreover,job failures occur in the cluster temporally but not evenly distributed in the cluster.Fourth,the mean time between failures of the cluster is calculated according to instance failure data,and the results show that most of the cluster nodes have the mean time between failures values as 1000 seconds,while a few of them have the mean time between failures values as 10000 seconds.
Research on Multi-recommendation Fusion Algorithm of Online Shopping Platform
ZHU Yu-jie, LIU Hu-chen
Computer Science. 2021, 48 (11A): 232-235.  doi:10.11896/jsjkx.201200010
Abstract PDF(1553KB) ( 832 )   
References | Related Articles | Metrics
The recommender system can help users solve the problem of information overload effectively and has been widely applied in major online shopping platforms.For users,a good recommendation algorithm can help them find products which meet their needs from a large number of products.For merchants,timely presentation of appropriate items to users can help merchants achieve precision marketing,discover long-tail products and recommend them to users to increase sales.Collaborative filtering and content-based recommendation are currently mature recommendation methods,but these methods have problems such as data sparsity,cold start,poor scalability,and difficulty in extracting multimedia information features.Therefore,this paper proposes a personalized recommendation algorithm based on the fusion of LR-GBDT-XGBOOST,which can effectively alleviate the above problems.Experiments are carried out under the official dataset of the Alibaba Tianchi big data competition.The results show that the proposed algorithm reduces the recommended sparsity and improves the accuracy of the recommendation.
Community Mining Based on KL-Ball
LOU Zheng-zheng, WANG Guan-wei, LI Hui, WU Yun-peng
Computer Science. 2021, 48 (11A): 236-243.  doi:10.11896/jsjkx.210300205
Abstract PDF(2560KB) ( 478 )   
References | Related Articles | Metrics
This paper presents a new community mining method where each community is viewed as a KL-Ball,and the KL divergence is adopted to measure the distance between nodes in the sparse adjacency matrix.This paper defines the KL-Ball consisting of four aspects:center,radius,mutual information and density.The center determines the location of KL-Ball in the complex network,and the radius determines the region of KL-Ball.The mutual information is used to measure the consistency of objects in a community,and the density is adopted to measure the coherence of a community.Given a radius,we aim to find communities with lower-information and higher-density in the complex network.For this purpose,we define the community mining objective function based on KL-Ball.Then we propose an optimization algorithm to minimize the objective function and theoretically prove the convergence of it.The proposed algorithm adopts a flexible community mining framework,and can be applied to several kinds of community mining tasks based on different locations and regions of the KL-Balls,such as the traditional community mining,high precision community mining and overlapping community detection.The experiment results show that the proposed KL-Ball based method can effectively find the community structure in complex network,including non-overlapping and overlapping communities.
High-order Collaborative Filtering Recommendation System Based on Knowledge Graph Embedding
XU Bing, YI Pei-yu, WANG Jin-ce, PENG Jian
Computer Science. 2021, 48 (11A): 244-250.  doi:10.11896/jsjkx.210100211
Abstract PDF(3243KB) ( 823 )   
References | Related Articles | Metrics
For the problem of data sparsity in recommendation systems,traditional collaborative filtering methods fail to capture the correlations between auxiliary information,which decreases the accuracy of recommendation.In this paper,we propose the KGE-CF model by introducing the knowledge graph as auxiliary information,and utilizing the multi-source structured data in knowledge graph to effectively alleviate data sparsity.KGE-CF integrates multi-layer perceptron to capture the high-order nonli-near features to learn deeper interaction information between users and items,which can improve the quality of recommendation.Specifically,we first map the user's historical interaction items with the corresponding entities in the knowledge graph,and use the translation model of the knowledge graph for training,through that we can get the entity embedding vector and the relation embedding vector,moreover the model can learn a richer user vector by interest disseminating.Then,after concatenating the obtained user vector and the item vector,we send it into the multi-layer perceptron to capture the high-order feature information between the user and the item.Finally,a sigmoid function is used to obtain the user's preference probability for candidate items.The experimental results on the real-world datasets prove that the proposed KGE-CF model in this paper has achieved the best recommendation performance than state-of-art methods.
User Interest Dictionary and LSTM Based Method for Personalized Emotion Classification
WANG You-wei, ZHU Chen, ZHU Jian-ming, LI Yang, FENG Li-zhou, LIU Jiang-chun
Computer Science. 2021, 48 (11A): 251-257.  doi:10.11896/jsjkx.201200202
Abstract PDF(1914KB) ( 742 )   
References | Related Articles | Metrics
Microblog is a social platform that people can share life,express opinions and vent emotions.Due to the large amount of data and easy access,the Microblog data has been widely used in emotion prediction for the web users.The traditional research on emotion classification of Microblog simply stays on the meaning of words,without considering the influence from the individuation of each person's language preference and style,which results a lower accuracy of the emotion classification.Firstly,this paper constructs a user interest dictionary by analyzing user interest characteristics and proposes a user interest dictionary basedemotion classification model.Secondly,by using the advantage of high classification accuracy of Long Short-Term Memory (LSTM),this paper trains a common LSTM based classification model.Finally,this paper fuses different models by using Support Vector Machine to obtain the final emotion classification results.The experimental results show that,compared with traditional classifiers such as SVM and Naive Bayesian,the personalized emotion classification method based on user interest dictionary and LSTM has a great improvement on classification accuracy.Compared with typical deep learning methods like LSTM andRecurrent Neural Network,the proposed method can obtain higher classification accuracy while ensuring the execution efficiency.
Research on Creep Feature Extraction and Early Warning Algorithm Based on Satellite MonitoringSpatial-Temporal Big Data
LIU Ya-chen, HUANG Xue-ying
Computer Science. 2021, 48 (11A): 258-264.  doi:10.11896/jsjkx.201000071
Abstract PDF(4414KB) ( 556 )   
References | Related Articles | Metrics
This paper presents a new methodology for overcoming the difficulties to issue warning of the occurrence time and tendency of the geologic hazards timely and accurately,employs the latest BeiDou satellite deformation monitoring technology,and considers a novel feature extraction of creep deformation method and algorithm for hazards warning.Based on the data analysis and data cleansing of the data from satellite monitoring spatial-temporal big data,we focus on time and spatial attribute and variation between different monitoring points.Moreover,we extract multi-dimensional of creep deformation such as displacement,displacement angle,instantaneous velocity,acceleration,etc.And the internal variation trend of the monitored data is displayed in a multi-dimensional manner.Research on creep disaster warning algorithm can find and warn potential disasters in deformation process,this finding is helpful to take measures to ensure the personal and property safety timely.Our research findings have important application value and theoretical significance in many fields.
Image Processing & Multimedia Technology
Light-weight Object Detection Network Optimized Based on YOLO Family
XU Yu-jun, LI Chen
Computer Science. 2021, 48 (11A): 265-269.  doi:10.11896/jsjkx.201000152
Abstract PDF(2226KB) ( 1045 )   
References | Related Articles | Metrics
Object detection is an active research field in the computer vision field.It is a very effective method to improve object detection precision by designing a large-scale deep convolutional neural network.However,it is unfavorable to deploy a large-scale object detection network in memory-limited applications.To solve the above problems,this paper proposes a light-weight object detection network which is based on design principles from the YOLO family of single-shot object detection network architectures.This network integrates the Ghost Module in GhostNet,in addition,a better Efficient Channel Attention (ECA) module is added to the convolution block by referring to the Squeeze-and-Excitation (SE) module in MobileNet-v3.This module can make better use of the available network capacity,making the network achieve a strong balance between reducing the complexity of architecture and computation and improving the performance of the model.In addition,Distance-IoU loss is used to solve the problem of inaccurate regression position of bounding box and effectively speeds up network convergence.Finally,the number of parameters of the model was compressed to 1.54 MB less than YOLO Nano (4.0MB),and the mAP on the VOC2007 data set was 72.1% higher than the existing YOLO Nano (69.1%).
Voiceprint Recognition Based on LSTM Neural Network
LIU Xiao-xuan, JI Yi, LIU Chun-ping
Computer Science. 2021, 48 (11A): 270-274.  doi:10.11896/jsjkx.210400041
Abstract PDF(2406KB) ( 1161 )   
References | Related Articles | Metrics
Voiceprint recognition determines the identification of the given speaker by voice,using the individual differences of biological characteristics.It has a wide range of use,with the characteristics of non-contact,simple acquisition,feature stability and so on.The existing statistical methods of voiceprint recognition have the limitations of single-source extracted feature and weak generalization ability.In recent years,with the rapid development of artificial intelligence and deep learning,neural networks are emerging in the field of voiceprint recognition.In this paper,a method based on Long Short-Term Memory (Long Short-Term Memory,LSTM) neural network was proposed to realize text-independent voiceprint recognition,using spectrograms to extract voiceprint features as the model input.Spectrograms can represent the frequency and energy information of voice signal in time direction comprehensively,and express more abundant voiceprint features.LSTM neural network is good at capturing temporal features,focusing on the information in time dimension,which is more consistent with the characteristics of voice data compared with other neural network models.The method in this paper combined the long-term learning of LSTM neural network with the sequential feature of voiceprint spectrograms effectively.The experimental results show that 84.31% accuracy is achieved on THCHS-30 voice data set.For three seconds short voice in natural environment,the accuracy of this method is 96.67%,which is better than the existing methods such as Gaussian Mixture Model and Convolutional Neural Network.
Complete Graph Face Clustering Based on Graph Convolution Network
WANG Wen-bo, LUO Heng-li
Computer Science. 2021, 48 (11A): 275-277.  doi:10.11896/jsjkx.201200102
Abstract PDF(1708KB) ( 671 )   
References | Related Articles | Metrics
Face clustering is a method of grouping face images according to different identities,which is mainly used in the fields of face annotation,image management.etc.There is massive redundant data in existing methods.To handle this issue,this paper uses a link prediction method based on complete graph constraint and context relationship.The clustering algorithm is based on graph convolution network for link prediction,combined with complete graph constraints to filter data,and the link relationship is constantly updated in the process of prediction.Experimental results show that the face clustering method combined with complete graph constraint can reduce redundant data,speed up the operation,and improve the accuracy of clustering.Thus it improves the overall performance of clustering.
Low-light Image Enhancement Method Based on U-net++ Network
LI Hua-ji, CHENG Jiang-hua, LIU Tong, CHENG Bang, ZHAO Kang-cheng
Computer Science. 2021, 48 (11A): 278-282.  doi:10.11896/jsjkx.210300111
Abstract PDF(3715KB) ( 879 )   
References | Related Articles | Metrics
Low-light image enhancement is one of the most challenging tasks in computer vision.The current algorithms have some problems,such as uneven brightness,low contrast,color distortion and serious noise.In this paper,a more natural dark light enhanced network framework based on improved U-net++ network is proposed.First of all,the low light image is input to the improved U-net++ network,and the dense connection of each layer is used to enhance the correlation of different levels of image features.Secondly,the image features of each level are fused and input to the convolution network layer for detail reconstruction.The experimental results show that the proposed method not only improves the brightness of the image,but also restores the detail features of the low light image better,and the color feature of the normal light image is closer to the nature.Tests on the PASCAL VOC test set show that the two important indicators,structural similarity (SSIM) and peak signal-to-noise ratio (PSNR),are 0.87 and 26.36,which are 18.6% and 11.4% higher than similar optimal algorithms respectively.
Road Surface Object Detection from Mobile Phone Based Sensor Trajectories
JIAO Dong-lai, WANG Hao-xiang, LYU Hai-yang, XU Ke
Computer Science. 2021, 48 (11A): 283-289.  doi:10.11896/jsjkx.210200145
Abstract PDF(5354KB) ( 699 )   
References | Related Articles | Metrics
Aimed at the problem of low efficiency and high cost in the traditional road surface object collection procedure,the method of road surface object recognition from mobile phone based sensor trajectories is proposed.Mobile phones are used to record the data changes of various sensors in the process of driving,and then the acceleration data after attitude correction are analyzed to find the relationship between the acceleration trend and the road condition.Finally,the constructe the BP neural network model,and use the acquired data to train the BP neural network model to recognize the road surface object and its position.Experiment results show that,the road surface object can be fast and accurately recognized by the mobile phone based senor trajectories,and the accuracy can be higher than 85%.In this paper,attitude of mobile acceleration sensor has carried on the real time correction.Because the acceleration changing of the Mobile phones is perpendicular to the road,we use the acceleration change to detecte the road feature.The method has nothing to do with a mobile phone accelerometer gesture,in addition,hardware requirements of the method are low,theefficiency of data acquisition is high,which reduce the cost of the road surface features information acquisition.
Research on Rockfall Detection Method of Mountain Railway Slope Based on YOLOv3 Algorithm
LIU Lin-ya, WU Song-ying, ZUO Zhi-yuan, CAO Zi-wen
Computer Science. 2021, 48 (11A): 290-294.  doi:10.11896/jsjkx.201200113
Abstract PDF(3881KB) ( 1341 )   
References | Related Articles | Metrics
The existing detection methods have the disadvantages of high detection cost and complex operation.In view of the above problems,this paper proposes to use smart phones and civilian cameras combined with light compensation device to collect various rock samples of various sizes and shapes in mountainous areas,and use deep convolution network to learn and extract the corresponding characteristics of rock samples for training.At the same time,yolov3 algorithm is introduced to build the depth learning model of slope rockfall detection of mountain railway,so as to realize the real-time detection of slope rockfall along the mountain railway.In addition,fast RCNN algorithm is set as a parallel comparative experiment.The experimental results show that the two detection algorithms can achieve high detection accuracy.Compared with fast RCNN algorithm,the detection accuracy of yolov3 algorithm is relatively low,but it is more sensitive to the small rockfall target,more capturing,and the detection speed is faster,which can better meet the actual engineering needs.
Low-quality Video Face Recognition Method Based on Super-resolution Reconstruction
LU Yao-yao, YUAN Jia-bin, HE Shan, WANG Tian-xing
Computer Science. 2021, 48 (11A): 295-302.  doi:10.11896/jsjkx.201200159
Abstract PDF(2420KB) ( 867 )   
References | Related Articles | Metrics
With the rise of deep neural networks,face recognition technology has developed rapidly.However,S2V (Still to Video)face recognition for low-quality video that is poor lighting conditions and low resolution still does not achieve the expected results,because the heterogeneous matching problem between the test video of low-quality and the high-definition image of the sample library.To solve this problem,this paper proposes a face recognition method for low-quality video based on super-resolution reconstruction.First,it uses clustering algorithm and random algorithm to select key frames for low-quality video frames based on face pose.Then,it builds a super-resolution reconstruction model S2V-SR for low-quality video S2V face recognition,and performs super-resolution reconstruction on key frames to obtain super-resolution key frames with higher resolution and more identity features.Finally,it uses the video face recognition network to extract deep features for classification and voting to obtain the final result.The proposed method is experimentally tested on the COX video face data set,and the best recognition accuracy is 55.91% and 70.85% in the relatively high-quality cam1 and cam3 videos,while in the relatively low-quality cam2 video the re-cognition accuracy rate second only to the best method is obtained.Experiments show that the proposed method can solve the hete-rogeneous matching problem in S2V face recognition to a certain extent,and obtain higher recognition accuracy and stability.
Hyperspectral Image Denoising Based on Robust Low Rank Tensor Restoration
WU Yong, LIU Yong-jian, TANG Tang, WANG Hong-lin, ZHENG Jian-cheng
Computer Science. 2021, 48 (11A): 303-307.  doi:10.11896/jsjkx.210200103
Abstract PDF(4436KB) ( 777 )   
References | Related Articles | Metrics
Denoising is an important preprocessing step to further analyze the hyperspectral image (HSI),and many denoising methods have been used for the denoising of the HSI data cube.However,the traditional denoising methods are sensitive to outliers and non-Gaussian noise.In this paper,by making using of the low-rank tensor property of the clean HSI data and the sparsity property of the outliers and non-Gaussian noise,we propose a new model based on the robust low-rank tensor recovery,which can retain the global structure of HSI and clean the outliers and mixed noise.The proposed model can be solved by the inexact augmented Lagrangian method.Experiments on simulated and real hyperspectral data show that the proposed algorithm is efficient for HSI restoration.
Scar Area Calculation Based on 3D Image
YAO Nan, ZHANG Zheng
Computer Science. 2021, 48 (11A): 308-313.  doi:10.11896/jsjkx.201100044
Abstract PDF(4038KB) ( 802 )   
References | Related Articles | Metrics
At present,forensic medicine mainly uses artificial methods to identify the area of injured scars,which has some instability and time-consuming problems.Therefore,a forensic identification method based on the 3D image to calculate the scar area is proposed.Firstly,a 3D laser scanner is used to obtain 3D image data of the unidentified scar.Secondly,the data is preprocessed to remove the background and noise,and the point cloud resolution is adjusted by down-sampling.Then the color area growth method is used to perform automatic region segmentation,also manual interaction is supplemented to adjust the target scar area.Finally,after surface reconstruction,the target scar is used to calculate its area.The results show that,compared with the current digital processing method of forensic medicine,the error is kept within 5% and the time consumption is reduced by more than 20%.
Adaptive Window Binocular Stereo Matching Algorithm Based on Image Segmentation
CAO Lin, YU Wei-wei
Computer Science. 2021, 48 (11A): 314-318.  doi:10.11896/jsjkx.201200264
Abstract PDF(3756KB) ( 632 )   
References | Related Articles | Metrics
Aiming at the problem that the traditional binocular stereo matching algorithm uses fixed window,which leads to low matching accuracy in weak texture regions,an adaptive window stereo matching algorithm based on image segmentation is proposed.Firstly,the mean shift algorithm is used to segment the image,and then the gray standard deviation of local sub regions is calculated.Based on this,an adaptive window size setting operator is proposed according to the texture richness.Based on the adaptive window size setting,the matching cost is calculated by combining census transform and gradient value,and the initial disparity is calculated by adaptive weight cost aggregation and “winner takeall” strategy respectively.Finally,the dense disparity map is obtained by using the principle of left and right disparity consistency and weighted median filtering.The adaptive window matching algorithm and fixed window matching algorithm proposed in this paper are used to match standard images on Middlebury dataset.The experimental results show that the average matching error rate of the proposed algorithm is 2.04%,which is 4.5% and 7.9% lower than that of the contrast algorithm.
Acoustic Emission Signal Recognition Based on Long Short Time Memory Neural Network
ZHOU Jun, YIN Yue, XIA Bin
Computer Science. 2021, 48 (11A): 319-326.  doi:10.11896/jsjkx.210700034
Abstract PDF(4587KB) ( 997 )   
References | Related Articles | Metrics
Acoustic emission testing does not need to enter the tested object for testing,and compared with traditional nondestructive testing technology,acoustic emission testing has unique advantages such as real-time,integrity and high sensitivity.Parameter analysis and wavelet analysis methods was used in the acoustic emission signal feature extraction,which lack of theoretical guidance and certain subjectivity.The BP neural network training easy to fall into local extremum when used in the acoustic emission signal recognition,Long Short-Term Memory neural network can learn step by step input sequence data and adaptive feature extraction,avoids artificial selection and extraction of the feature.Z-score standardization was used in the acoustic emission signal,LSTM neural network uses single hidden layer structure,compare the different correct recognition rate on the test set in different situations of learning algorithm,the number of hidden layer neurons,the output neurons dropout rate,optimal model of acoustic emission signal recognition based on LSTM structure was found out.Comparing the acoustic emission signal recognition accuracy and with the BP neural network,experimental results show that the recognition rate of LSTM neural network is 76.51% under the setting of Adam algorithm,250 hidden layer neurons and 0.5 dropout rate,Compared with the highest recognition rate of 53.9% of BP neural network,the proposed algorithm has advantages.
Nighttime Image Dehazing Method Based on Adaptive Light Source Region
WANG Tong-sen, SHI Qin-zhong, WANG De-fa, DONG Shuo, YANG Guo-wei, YU Teng
Computer Science. 2021, 48 (11A): 327-333.  doi:10.11896/jsjkx.210300072
Abstract PDF(4975KB) ( 744 )   
References | Related Articles | Metrics
Nighttime hazy image will cause image quality degradation,mainly reflected in the uneven illumination,low contrast and serious color deviation of nighttime hazy image,and artificial light source makes the ambient light nonuniformity.The existing mainstream algorithms are mainly for daytime image processing,but are not suitable for night scene dehazing processing.This makes night dehazing more difficult.In order to solve the above problems,this paper analyzes the imaging features of night image with fog and proposes a new night image dehazing algorithm.Aiming at the problem of color deviation of hazy images at night,this paper proposes an improved maximum reflectance prior algorithm (MRP) for color correction.This method operates each color channel separately for color correction,so as to reduce the halo effect around the light source area caused by MRP.As for the characteristics of nonuniformity of ambient light in night scene,a minimum reflectance prior algorithm based on low frequency component of hazy images is proposed.In order to solve the problem that the dark channel prior (DCP) estimation of transmittance fails at the light source,we propose a region adaptive algorithm of transmittance estimation based on the light source.The experimental results show that the proposed algorithm can suppress the halo and the divergence of the light source area.At the same time,it can better reproduce the dark details and restore the image with better brightness.
SCTD 1.0:Sonar Common Target Detection Dataset
ZHOU Yan, CHEN Shao-chang, WU Ke, NING Ming-qiang, CHEN Hong-kun, ZHANG Peng
Computer Science. 2021, 48 (11A): 334-339.  doi:10.11896/jsjkx.210100138
Abstract PDF(4260KB) ( 1676 )   
References | Related Articles | Metrics
In recent years,convolutional neural networks (CNN) have been widely used in large-scale natural image datasets (such as ImageNet,COCO).However,there is a lack of applied research in the field of sonar image detection and recognition,which suffers from a lack of sonar image target detection and classification datasets and often faces sparse and unbalanced samples of underwater targets.In response to this problem,based on the extensive collection of sonar images,this paper constructs a completely open sonar common target detection dataset SCTD1.0 that can be used for sonar image detection and classification research.The dataset currently contains three types of typical targets:underwater shipwreck,wreckage of crashed aircraft,and victims,with a total of 596 samples.On the basis of SCTD1.0,this paper uses transfer learning to test the benchmarks of detection and classification.Specifically,for the detection task,the feature pyramid network is used to combine and utilize multi-scale features,and the performance of the three detection frameworks YOLOv3,Faster R-CNN,and Cascade R-CNN on this dataset is compared.For classification tasks,the classification performance of the three networks of VGGNet,ResNet50,and DenseNet is compared,and the classification accuracy rate reaches about 90%.
Centroid Method Based Target Tracking and Application for Internet of Vehicles
YE Yang, LU Qi, CHENG Shi-wei
Computer Science. 2021, 48 (11A): 340-344.  doi:10.11896/jsjkx.210200004
Abstract PDF(2693KB) ( 698 )   
References | Related Articles | Metrics
Vehicle target tracking is an indispensable part of the realization of the Internet of Vehicles,which aims to obtain vehicle dynamic information to improve the efficiency of traffic operation.Its core is to analyze and process the video images collected by a large number of monitoring probes to realize real-time detection and tracking of vehicles.In order to further improve the efficiency of target detection and reduce hardware costs,this paper proposes a foreground detection method based on the two-frame difference method,and a vehicle contour detection and tracking method based on the centroid method.Based on OpenCV3.4.1 and VS2017,the algorithm verification and simulation test are carried out.The results show that the accuracy of the algorithm for vehicle tracking reaches 92.3%,and the average processing time is 42.63 ms.It can be deployed and applied on embedded devices in the Internet of Vehicles.
Speed Limit Sign Recognition Based on LeNet-5 CNN and Color Feature
WANG Ji-min, WEI Yi, ZHOU Yu, SUN Ao, LIU Yuan-sheng
Computer Science. 2021, 48 (11A): 345-350.  doi:10.11896/jsjkx.201200213
Abstract PDF(4992KB) ( 804 )   
References | Related Articles | Metrics
Speed limit sign recognition is an important part of intelligent driving.This research analyzes the problems of existing methods.In order to improve the versatility and accuracy of neural networks on Chinese speed limit signs,in the detection part of speed limit signs,a new screening method based on color space is proposed.In the recognition part of the speed limit sign,the neural network is improved on the basis of the existing LeNet-5 architecture.By fusing the German traffic sign data set (GTSRB) and Tsinghua traffic sign data set (TT100K),a new data set is made and sent to the neural network to train the model after data amplification.Using swish activation function,the optimal recognition accuracy rate obtained on the test set is 99.62%,and the model has strong anti-interference ability and strong practical performance.
Local Weighted Representation Based Linear Regression Classifier and Face Recognition
YANG Zhang-jing, WANG Wen-bo, HUANG Pu, ZHANG Fan-long, WANG Xin
Computer Science. 2021, 48 (11A): 351-359.  doi:10.11896/jsjkx.210100173
Abstract PDF(5142KB) ( 557 )   
References | Related Articles | Metrics
Linear regression classifier (LRC) is an effective image classification algorithm.However,LRC does not pay attention to the local structure information of data and ignores the differences among samples within the class,and the performance may degrade when the facial images contain variations in expression,illumination,angle and occlusion.To address this problem,a linear regression classifier based on local weighted representation (LWR-LRC) is proposed.Firstly,LWR-LRC constructs a weighted representative sample for each class of samples based on the similarity between test samples and all samples,then decomposes the test samples into linear combinations of weighted representative samples,finally classifies the test samples into the category with the largest reconstruction coefficient.LWR-LRC considers the local structure of samples,constructs the optimal representative samples of each class of samples,and uses the representative samples to calculate,which improves the robustness and greatly time cost.The experiments on AR,CMU PIE,FERET and GT datasets show that LWR-LRC is superior to NNC,SRC,LRC,CRC,MRC and LMRC.
Multi-scale U Network Realizes Segmentation and Recognition of Tomato Leaf Disease
GU Xing-jian, ZHU Jian-feng, REN Shou-gang, XIONG Ying-jun, XU Huan-liang
Computer Science. 2021, 48 (11A): 360-366.  doi:10.11896/jsjkx.201000166
Abstract PDF(2998KB) ( 755 )   
References | Related Articles | Metrics
With the development of deep learning technology,convolutional neural network has been the mainstream method for plant leaf disease recognition and disease spot segmentation.Aiming at the problems of different sizes and irregular shapes of tomato leaf lesions,need for a large number of pixel-level labels,a novel multi-scale U network is proposed,which realizes tomato leaf lesion segmentation and disease recognition simultaneously.For disease feature extraction,a multi-scale residual module including different sizes of receptive fields is used to extract disease features according to the different disease spot size and shape.The CB module (Classifier and Bridge) is introduced to connect the disease feature extraction stage with the lesion segmentation stage,which classifies the disease and also reversely generates an activation map of specific class according to the classification result.This activation map contains the specific type of lesions label information.In the segmentation stage,upsampling and convolution are used to deconvolve the activation map.The deconvolution feature and the low-level feature are merged by the jump connection method.In order to make lesion location segmentation more accurate,a few of pixel-level labels are used for training to minimize two-class cross-entropy loss.In the experiment,the original samples and samples with simulated noise and light intensity are used to verify the performance of disease spot segmentation and disease recognition of our method.On the original sample set,the average pixel accuracy,average intersection ratio,and frequency weight intersection ratio of our method reaches 86.15%,75.25%,and 90.27%,respectively.In the interference sample with 30% brightness reduction,salt and pepper noise,the recognition accuracy of our method obtains 95.10% and 99.20% respectively.Experimental results show that the proposed method can achieve improvement in segmentation and recognition of tomato leaf lesions simultaneously.
Method for Diagnosis and Location of Chest X-ray Diseases with Deep Learning Based on Weak Supervision
Computer Science. 2021, 48 (11A): 367-369.  doi:10.11896/jsjkx.201200152
Abstract PDF(1791KB) ( 676 )   
References | Related Articles | Metrics
Chest disease is one of the most common diseases,chest X-ray is an important method for examination and diagnosis of chest disease.The construction of a highly accurate prediction model for chest diagnosis usually requires a large number of tags and manual annotation of abnormal locations.However,it is difficult to obtain such annotated data,especially data with location annotation.Therefore,building a method that use only a few location annotations becomes a major problem.Although there have been some weak supervision methods to solve this problem,most of them only focus on the image information,rarely considering the relationship between the images.Therefore,a chest X-ray disease diagnosis and location model based on weak supervised deep learning is proposed.While deep learning is used to extract the image information,the graph structure is introduced and hash code is used to add the similarity of the image itself and the relational information of the image into the learning process.In the case of no additional annotation,a small amount of annotation can achieve a good recognition and location effect.After validation on the Chest X-ray dataset,the location accuracy (IoU) is 44% when using only 3% of the location-tagged data.This indicates that this method can effectively identify and locate chest X-ray lesions,provide doctors with candidate areas for screening,and assist doctors in the diagnosis of chest diseases.
Multi-object Tracking Algorithm Based on YOLOv3 and Hierarchical Data Association
LIU Yan, QIN Pin-le, ZENG Jian-chao
Computer Science. 2021, 48 (11A): 370-375.  doi:10.11896/jsjkx.201000115
Abstract PDF(2632KB) ( 702 )   
References | Related Articles | Metrics
In order to alleviate the real-time problem of multi-object tracking methods and the tracking difficulty caused by the high similarity of appearance and the excessive number of error detection in the tracking process,a new multi-object tracking method is proposed,which is based on the improved YOLOv3 and hierarchical data association.As the lightweight network MobileNet uses the deep separable convolution to compress the original network,so as to reduce the network parameters,we uses MobileNet to replace the main structure of YOLOv3 network while retaining the multi-scale prediction part of YOLOv3,so as to reduce the complexity of the network and make the method meet the real-time requirements.Compared with the detection network used in other multi-object tracking methods,we proposed detection network model size is 91 M,and the single detection time can reach 3.12 s.At the same time,the algorithm introduces hierarchical data association method based on object appearance features and motion features.Compared with the method using only appearance features,the hierarchical data association method improves the evaluation index MOTA by 6.5 and MOTP by 1.7.On the MOT16 data set,the tracking accuracy can reach 77.2% and has good anti-jamming ability and real-time performance.
Semantic Segmentation of SAR Remote Sensing Image Based on U-Net Optimization
WANG Xin, ZHANG Hao-yu, LING Cheng
Computer Science. 2021, 48 (11A): 376-381.  doi:10.11896/jsjkx.210300260
Abstract PDF(2840KB) ( 1055 )   
References | Related Articles | Metrics
Multi-spectral image segmentation is an important basic link in remote sensing image interpretation.SAR remote sensing images contain complex object information.Traditional segmentation methods have problems such as time-consuming and low efficiency,which limits the application of traditional image segmentation methods.In recent years,the application of deep learning algorithms in the direction of computer vision has achieved good results.Aiming at the problem of semantic segmentation of multi-spectral remote sensing images,deep learning semantic segmentation methods are used to achieve high-performance segmentation of remote sensing images.In the U-Net network structure Above,add activation layer,dropout layer,convolutional la-yer,and propose a deep convolutional neural network optimized based on U-Net.On the basis of a small amount of data set,it realizes for rapid detection of buildings and rivers,the overall segmentation accuracy rate reaches 94.6%.The comparison test results with U-Net and SegNet show that the segmentation accuracy of the method used in this paper is better than that of U-Net and SegNet.Compared with U-Net and SegNet,it has increased by a minimum of 2.5% and 5.8%,respectively.
Feature Classification Method Based on Improved DeeplabV3+
ZHU Rong, YE Kuan, YANG Bo, XIE Huan, ZHAO Lei
Computer Science. 2021, 48 (11A): 382-385.  doi:10.11896/jsjkx.201100184
Abstract PDF(2451KB) ( 781 )   
References | Related Articles | Metrics
The original DeeplabV3+ algorithm is not accurate enough for the edge segmentation of UAV aerial images,and the road segmentation is discontinuous.Therefore,in order to solve these problems,this paper improves the DeeplabV3+ algorithm.Firstly,the feature fusion is carried out in the coding stage to enhance the semantic information of the shallow feature map.Secondly,the boundary extraction branch module is added to the segmentation network structure,and Canny edge detection algorithm is used to extract the real boundary information for supervision training,so that the network can segment the edge of ground objects.Finally,in the decoding stage of the network,more shallow features are fused.The experimental results show that the mIoU value of the proposed method is 80.92%,which is 6.35% higher than that of the DeeplabV3+ algorithm,and can effectively classify the ground objects.
Light Superposition-based Color Constancy Computational Method
FENG Yi-fan, ZHAO Xue-qing, SHI Xin, YANG Kun
Computer Science. 2021, 48 (11A): 386-390.  doi:10.11896/jsjkx.210200053
Abstract PDF(3008KB) ( 747 )   
References | Related Articles | Metrics
Color constancy is a psychological tendency of the human visual system the color perception of external visual stimuli.This cognitive function of human vision can adaptively ignore external changes,and perceive colors steadily.Inspired by the color perception of the human visual system,considering the capability that computer vision tasks eliminate the effect of external light automatically,it is of important research significance to restore the true color information of objects and provide stable color characteristics.This paper proposes a light superposition-based color constancy computational method,which can effectively eliminate the influence of changes in the spectral composition of external light on the color of objects.First of all,the MAX-MEAN method is proposed to estimate the illumination in the scene (MM estimation,in short),that is,estimating the illumination in the achromatic scene by the average reflection and maximum reflection of all object surfaces in the scene.Then,based on the MM estimation,the color constant calculation method of light superposition is used to obtain the final color-free image.11346 indoor and outdoor scene images in the SFU Gray-ball public data set are used for simulation and validation.The experimental results show that the illumination superposition color constant calculation method proposed in this paper can effectively estimate the illumination information,perform the color constancy calculation,and get the image without color constancy.
Clothing Image Sets Classification Based on Manifold Structure Neural Network
CHENG Ming, MA Pei, HE Ru-han
Computer Science. 2021, 48 (11A): 391-395.  doi:10.11896/jsjkx.201200127
Abstract PDF(2693KB) ( 779 )   
References | Related Articles | Metrics
Clothing classification based on deep learning has developed rapidly with the release of large-scale fashion data sets.However,most of the current clothing image classification methods are performed in a scene where the same clothing has a single,frontal or close-to-front image,which leads to misclassification of clothing when the view of clothing changes.In reality,the clothing features such as large deformation and severe occlusion further aggravate the problem.Therefore,a clothing image set classification method based on manifold structure neural network is proposed which uses manifold space to better represent the internal structure characteristics of clothing.Concretely,first,the shallow features of the clothing image set are extracted through the traditional convolutional neural network,and then the Euclidean feature data are converted into manifold data by using the covariance pooling.Finally,the internal manifold structures of clothing image sets are learned through the neural network based on manifold structure to obtain more accurate classification results.The experimental results show that the Precision,Recall and F-1 score of the proposed method on the MVC dataset can reach 89.64%,89.12% and 88.69%.Compared with the existing image sets (video) classification algorithms,the proposed method obtains an improvement of 2.04%,2.65% and 2.70%.It is illustrated that the proposed method is more accurate,efficient and robust than existing methods.
Study on Digital Tube Image Reading Combining Improved Threading Method with HOG+SVM Method
SONG Yi-yan, TANG Dong-lin, WU Xu-long, ZHOU Li, QIN Bei-xuan
Computer Science. 2021, 48 (11A): 396-399.  doi:10.11896/jsjkx.210100123
Abstract PDF(2695KB) ( 613 )   
References | Related Articles | Metrics
In traditional projection method when rely too much on a single digital image are extracted image binarization and tilt correction effect problem,using a method based on contour extraction and contour sort of digital image segmentation method,experimental results show this method is compared with the projection segmentation on the success rate for segmenting the digital area increased by 13.5%;Against traditional threading method for number 1 low recognition and machine learning algorithms run takes longer problem,put forward an improved,based on the six characteristics of segment digital tube threading method and the HOG+SVM method with the combination of digital identification method,the method of digital tube digital identification accuracy than traditional threading method by about 4.5%,and the average elapsed time only 1/5 ofthe HOG+SVM method.The experimental results show the reliability and effectiveness of the method in digital tube reading.
Visual Human Action Recognition Based on Deep Belief Network
HONG Yao-qiu
Computer Science. 2021, 48 (11A): 400-403.  doi:10.11896/jsjkx.210200079
Abstract PDF(2800KB) ( 533 )   
References | Related Articles | Metrics
In order to solve the problem of human action recognition in a large number of videos with complex backgrounds and changing viewpoints on the Internet,an innovative method for human action recognition using unsupervised deep belief networks (DBNs) is proposed.This method uses deep belief networks (DBNs) and restricted Boltzmann machines for unconstrained video action recognition,and uses an unsupervised deep learning model to automatically extract appropriate feature representations without any prior knowledge.Through the identification of a challenging UCF sports data set,the method is proved to be accurate and effective.At the same time,this method is suitable for other visual recognition tasks,and will be extended to unstructured human activity recognition in the future.
Video Synopsis Based on Trajectory Spatial Relationship Analysis
QU Zhi-guo, TAN Xian-si, TANG Tang, ZHENG Jian-cheng, FEI Tai-yong
Computer Science. 2021, 48 (11A): 404-408.  doi:10.11896/jsjkx.210100125
Abstract PDF(3738KB) ( 539 )   
References | Related Articles | Metrics
Collision phenomenon is an unpleasant issue that needs to be addressed during trajectory rearrangement in video synopsis.It is usually constrained by some collision cost in the final energy function to be optimized.However,most synopsis methods compute the collision cost term repeatedly in the iterative optimization process,leading to serious computation redundancy.To solve that,a novel synopsis method based on spatial relationships between trajectories is proposed in this paper.It turns out whether two trajectories will collide or not can be determined beforehand by analyzing their spatial relationships.Accordingly,three kinds of relationship are defined and corresponding fast computation of collision cost are given.In this way,the redundancy in collision cost computation is decreased and thus improving the speed of traditional methods obviously.Experimental results demonstrate the effectiveness of the proposed method.
Face Anti-spoofing Algorithm Based on Multi-feature Fusion
LUAN Xiao, LI Xiao-shuang
Computer Science. 2021, 48 (11A): 409-415.  doi:10.11896/jsjkx.210100181
Abstract PDF(2282KB) ( 1227 )   
References | Related Articles | Metrics
In recent years,with the development of face recognition systems,various spoofing methods that impersonate legitimate users appear.Face anti-spoofing detection method based on a single clue no longer meets the requirements of current face recognition system under complex environment.Based on this,we propose to use a convolutional neural network to learn multi-feature from different clues of face images,and to fuse the depth map,the face optical flow map,and the residual noise map to perform liveness detection.The depth map can distinguish the depth information between real and fake faces in space,the optical flow map can distinguish the dynamic information between real and fake faces in time,the residual noise map is based on the one-time imaging of the real face and the fake face.The secondary imaging noise components are distinguished by different components,and the three features are merged to use space,time and multi-dimensional clues to make up for the shortcomings of a single clue,and also improve the generalization ability of the model.Compared with the existing methods,our method shows promising results both on the single database and cross-databases.Specifically,equal error rate (EER) of our method on databases of CASIA,REPLAY-ATTACK and NUAA can achieve 0.11%,0.06% and 0.45%,respectively.
Lightweight Lane Detection Model Based on Row-column Decoupled Sampling
CHEN Hao-nan, LEI Yin-jie, WANG Hao
Computer Science. 2021, 48 (11A): 416-419.  doi:10.11896/jsjkx.201100206
Abstract PDF(2142KB) ( 662 )   
References | Related Articles | Metrics
With the development of deep learning,lane detection model based on deepconvolution neural network has been widely applied in autonomous driving system and advanced driving assistant system.These models achieve high accuracy but usually have the disadvantages of large computation and high latency.In order to solve this problem,a specially designed lightweight network for lane detection is proposed.Firstly,a convolution module with row-column decoupled sampling is proposed,which optimizes traditional residual convolution module by utilizing the row-column decomposability of lane area in the image.Secondly,the depth-wise separable convolution technology is used to further reduce the computational complexity of the row-column decoupled sampling convolution module.In addition,a pyramid dilation convolution module is designed to increase the receptive field of the mo-del.The experimental results on CULane dataset show that comparing with the state of the art model SCNN,the floating-point ope-rations of our model is reduced by 95.2% and F1-score is increased by 1.0%.The computation cost of lane detection model is significantly reduced while maintaining high accuracy.
Visibility Detection of Single Fogging Image Based on Transmittance and Scene Depth
ZHANG Ding, JIANG Mu-rong, HUANG Ya-qun
Computer Science. 2021, 48 (11A): 420-423.  doi:10.11896/jsjkx.210200072
Abstract PDF(2278KB) ( 861 )   
References | Related Articles | Metrics
The traditional methods of single image visibility detection have the problems of high hardware cost,small application range and low detection efficiency.This paper proposes a new method of single image scene depth visibility detection.Firstly,the visibility detection principle is deduced according to Koschmieder law and ICAO recommended contrast threshold,and then the extinction coefficient is obtained according to the atmospheric attenuation model,the key factors affecting the visibility of the image,such as the transmittance and the depth of the scene,are determined.Then the transmittance value is obtained by using the dark channel prior theory,and the depth value of the scene is obtained by combining the SFS (shape from shadow) and the binocular model.Finally,the visibility of the image is inversed by solving the extinction coefficient.The experimental results verify the effectiveness of the method,and the accuracy and detection efficiency are greatly improved compared with the traditional detection methods.This method does not need the internal parameters of the camera,and does not need to take multiple images of the same scene,so it is easy to operate and has a wide range of applications.
Classification Research of Remote Sensing Image Based on Super Resolution Reconstruction
XIE Hai-ping, LI Gao-yuan, YANG Hai-tao, ZHAO Hong-li
Computer Science. 2021, 48 (11A): 424-428.  doi:10.11896/jsjkx.210300132
Abstract PDF(2421KB) ( 737 )   
References | Related Articles | Metrics
Usually,super-resolution technology makes the image get better visual quality,which is conducive to the visual interpretation of the image.However,will super-resolution technology improve the effect of image in the application of classification,recognition and other more advanced computer vision tasks? Combined with the development of computer vision technology,a new image classification model is trained to classify the images processed by different super-resolution methods.By measuring the classification accuracy of different types of images,the promotion effect of different super-resolution methods on image classification task is measured.Experimental results on remote sensing classification data set show that the well-designed super-resolution model has obvious effect on enhancing low-quality images.Compared with the interpolation method,the image processed by super-resolution achieves higher classification accuracy in the classification model,which proves that the well-designed super-resolution model can promote the image classification task.It is confirmed that super-resolution can effectively improve the perception effect of image in computer vision.
Credit Risk Assessment Method of P2P Online Loan Borrowers Based on Deep Forest
WANG Xiao-xiao, WANG Ting-wen, MA Yu-ling, FAN Jia-yi, CUI Chao-ran
Computer Science. 2021, 48 (11A): 429-434.  doi:10.11896/jsjkx.201000013
Abstract PDF(2253KB) ( 662 )   
References | Related Articles | Metrics
P2P online lending is an emerging financial business model in recent years,which has many advantages of low investment threshold,convenient transaction and low financing cost.However,at the same time of rapid growth,the credit risk problem in the lending process has become increasingly prominent,and the endless stream of borrowers running away and even fraud have left a heavy shadow on the industry.Aiming at this problem,a credit risk assessment method of P2P online loan borrowers based on deep forest is proposed.Firstly,the features are extracted from the basic information and the historical loan information of the borrower.Then,the deep forest model was constructed by multi-granularity scanning and cascade forest module to predict the default of borrowers.At the same time,Gini index was used to calculate the feature importance score of random forest,and Borda count method was used to sort and fusion,so as to give a certain explanation to the prediction results of the model.On the two public datasets of LendingClub and Paipaidai,the proposed method was compared with methods such as support vector machines,random forests,and wide and deep networks.Experiments show that the method has better performance,and the feature importance rating is consistent with people's intuitive understanding and objective understanding.
Network & Communication
Semi-online Algorithms for Mixed Ring with Two Nodes
XIAO Man, LI Wei-dong
Computer Science. 2021, 48 (11A): 441-445.  doi:10.11896/jsjkx.201100153
Abstract PDF(1766KB) ( 474 )   
References | Related Articles | Metrics
Two semi-online scenarios of load balancing problem on a two-point mixed ring are studied.This paper gives a two-point mixed ring and several flow requests,finds a suitable flow transportation method to make the maximum load on the ring as small as possible.When there is a buffer with a capacity of K,the lower bound of 4/3 is given.In particular,when K=1,a lower bound of 3/2 is given,and a semi-online algorithm with a competitive ratio of at most 8/5 is given.When the sum of all flow demands is known,an optimal semi-online algorithm with a competitive ratio of 3/2 is designed.
Real Time Wireless Connection Scheme for Multi-nodes
QIAN Guang-ming, YI Chao
Computer Science. 2021, 48 (11A): 446-451.  doi:10.11896/jsjkx.201200209
Abstract PDF(1961KB) ( 564 )   
References | Related Articles | Metrics
A wireless connection scheme,named six synchronous successive transmission (SSST),is presented.The network using this scheme is called SSST network.It is mainly used for the real time radio scenario where multiple slave nodes request to send data to a master node.It has the same working frequency band with the Bluetooh.Based on SSST,slave nodes are allowed to send data transmission requests after synchronous packets are received from the master only.Slaves send their requests in an arranged way.The master transmits six synchronous packets continuously,each of which has a different index.The SSST protocol is collision free when multiple slaves require connecting the master simultaneously.The slave with the higher priority can be selected to access the network first,and the response time of each consecutive stage is easy to predict.All these are just the important features that real-time applications should have.The related theories are demonstrated and verified with experiments,and compared with the advertising and scanning of the Bluetooth.
Security Clustering Strategy Based on Particle Swarm Optimization Algorithm in Wireless Sensor Network
JIANG Jian-feng, SUN Jin-xia, YOU Lan-tao
Computer Science. 2021, 48 (11A): 452-455.  doi:10.11896/jsjkx.210900131
Abstract PDF(2435KB) ( 535 )   
References | Related Articles | Metrics
In order to solve the problem of short network survival time and lack of effective secure transmission mechanism in wireless sensor networks,a secure clustering strategy for wireless sensor networks based on particle swarm optimization algorithm(SC_PSO) is proposed.This strategy combines the polynomial hybrid key distribution technology to encrypt the communication data of nodes between clusters,which ensures the security performance of data transmission.On the other hand,an optimized particle swarm algorithm is used to construct a fitness function based on the remaining energy of the node and the communication distance to select the optimal cluster head and the number of clusters.It can solve the problem of energy loss caused by the encryption algorithm,and it can ensure the performance of the sensor network while realizing data security communication.The network simulation test shows that this strategy can increase the network throughput by 120% and extend the life cycle of the sensor network by 30%~65% while ensuring the security of the sensor network.
PSO-GA Based Approach to Multi-edge Load Balancing
YAO Ze-wei, LIU Jia-wen, HU Jun-qin, CHEN Xing
Computer Science. 2021, 48 (11A): 456-463.  doi:10.11896/jsjkx.210100191
Abstract PDF(2254KB) ( 558 )   
References | Related Articles | Metrics
As a new paradigm,mobile edge computing (MEC) can provide an efficient method to solve the computing and storage resource constraints of mobile devices.Through the wireless network,it migrates the intensive tasks on mobile devices to the edges near the users for execution,then the edges transmit the execution results back to mobile devices.Due to the randomness of users' movement,the load on each edge which deployed in the city is usually inconsistent.To solve the problem of multi-edge load balancing,the task scheduling is considered to minimize the maximum response time of tasks in the edge set,thereby improving the performance of mobile devices.Firstly,the multi-edge load balancing problem is formally defined.Then particle swarm optimization-genetic algorithm(PSO-GA) is proposed to solve the multi-edge load balancing problem.Finally,the performance of the algorithm is compared and analyzed with the random migration algorithm and the greedy algorithm through simulation experiments.The experimental results show that PSO-GA is superior to random migration and greedy algorithm by 51.58% and 26.34%,respectively.Therefore,PSO-GA has a better potential for reducing task response time of the edges and improving user experience.
Research on Two-level Scheduled In-band Full-duplex Media Access Control Mechanism
GUAN Zheng, LYU Wei, JIA Yao, YANG Zhi-jun
Computer Science. 2021, 48 (11A): 464-470.  doi:10.11896/jsjkx.201200026
Abstract PDF(3938KB) ( 460 )   
References | Related Articles | Metrics
In-band full-duplex (IBFD) wireless communication allows the nodes to send and receive simultaneously in the same frequency band,which is an effective way to improve the spectrum utilization rate.This paper proposes a two-level scheduled in-band full-duplex (TSIB-FD) for wireless networks,which divides the access process into three stages:information collection,full-duplex data transmission and acknowledgement.A first-level scheduling scheme is generated by AP in the information collection stage by node business requirements and inter-interference relations.During thefirst-level full-duplex link data transmission,AP can build the second-level scheduling dynamically according to current transmission.ACK is processed after complete bidirectional data transmission.The simulation results show that TSIB-FD has improved the system throughput with a lower delay.Compared with Janus,the throughput can be improved by 38.5%,and compared with HBPOLL,it can be improved by 100%.
Devices Low Energy Consumption Scheduling Algorithm Based on Dynamic Priority
ZHANG Yi-wen, LIN Ming-wei
Computer Science. 2021, 48 (11A): 471-475.  doi:10.11896/jsjkx.210100080
Abstract PDF(2052KB) ( 601 )   
References | Related Articles | Metrics
Previous studies consider independent periodic task model and only apply dynamic voltage frequency scaling (DVFS) to reduce energy consumption.An algorithm that can support preemptive periodic tasks with non-preemptive shared resources is proposed to overcome this shortcoming.It combines DVFS and dynamic power management (DPM) techniques to reduce energy consumption.It consists of device scheduling and job scheduling.In device scheduling,DPM technique is used to reduce the energy consumption of IO devices.In job scheduling,the earliest deadline first policy is used to schedule tasks and the stack resource protocol is used as synchronization protocol for shared resources.In addition,the task executes at low speed without blocking and switches to high speed with blocking to reduce the energy consumption of the processor.The experimental result shows that the proposed algorithm can yield significantly energy savings with respect to the existing algorithm.
Link Mapping Algorithm Based on Weighted Graph
GAO Ming, ZHOU Hui-ying, JIAO Hai, YING Li-li
Computer Science. 2021, 48 (11A): 476-480.  doi:10.11896/jsjkx.201200216
Abstract PDF(2327KB) ( 493 )   
References | Related Articles | Metrics
Service function chain (SFC),as a concept of service deployment,provides a higher flexibility for network.This paper studies the mapping problem in the service function deployment,and proposes a link mapping algorithm based on weighted graph for the service function chain's business choreography plane deployment,so as to balance the load requirements of the functional service nodes deployed to the physical nodes.And this paper presents a mapping algorithm for virtual link of service function,that is,the combination of service function is first carried out,and then the actual link situation is modeled and analyzed,the initial va-lue is obtained by using efficiency matrix,and finally the former is corrected by using heuristic algorithm.Through the modeling analysis and comparison with Eigen decomposition of adjacency matrices (Eigen) of map matching strategy which reduces the link bandwidth demand,the algorithm can complete the service request under the condition of balanced link node load and link bandwidth,and in the growing service chain length and flow,the algorithm is more stable to changes in throughput and can reduce the cost of mapping for existing physical network.
Synchronization of Uncertain Complex Networks with Sampled-data and Input Saturation
ZHAO Man-yu, YE Jun
Computer Science. 2021, 48 (11A): 481-484.  doi:10.11896/jsjkx.210100063
Abstract PDF(1712KB) ( 490 )   
References | Related Articles | Metrics
In actual network systems,there are widespread external interference and parameter mutations and other uncertain phenomena,which will cause the system to fail to achieve synchronization,and even destroy the stability of the system.Therefore,it is important to study uncertain complex networks.This paper discusses the synchronization problem of nonlinear uncertain complex dynamical networks with sampled-data and input saturation.Firstly,we establish a nonlinear and uncertain complex network model.Secondly,a leader is introduced to design a sampled-data control protocol with input saturation.By constructing an appropriate time-dependent Lyapunov functional and applying the stability theory,integral inequality method and the linear matrix inequality method,it is proved that the nonlinear uncertain complex dynamical networks can achieve synchronization under certain conditions,that is every follower can track the leader,and the sufficient criteria for achieving synchronization of nonlinear uncertain complex networks is derived.Finally,simulation examples are provided to verify the effectiveness and validity of the theoretical results.
QoS Optimization of Data Center Network Based on MPLS-TE
JIANG Jian-feng, YOU Lan-tao
Computer Science. 2021, 48 (11A): 485-489.  doi:10.11896/jsjkx.210900190
Abstract PDF(3084KB) ( 556 )   
References | Related Articles | Metrics
Network congestion and delay can be overcome by making full use of path allocation and relocation routing traffic methods.Multi-protocol label switching is one of the solutions to solve the QoS limitation of date center network service quality in IP networks.It designs a QoS algorithm based on traffic engineering and path allocation in combination with the QoS model to ensure the quality of network service.The network test results show that the algorithm can improve date center network transmission efficiency,reduce delay and packet loss rate,especially for voice and video traffic,the network performance has been greatly improved.
TCAM Multi-field Rule Coding Technique Based on Hypercube
WANG Yun-xiao, ZHAO Li-na, MA Lin, LI Ning, LIU Zi-yan, ZHANG Jie
Computer Science. 2021, 48 (11A): 490-494.  doi:10.11896/jsjkx.201100161
Abstract PDF(2970KB) ( 590 )   
References | Related Articles | Metrics
With the development and popularization of the Internet,the network scale,bandwidth and network packet transmission speed are growing exponentially.The increasingly rapid growth of network users also puts increasing pressure on the Internet infrastructure.As a key link to improve the performance of link bandwidth,the improvement of packet classification processing speed plays a key role in the development of various application services in the high-speed network environment.The current message classification algorithm has the problems of insufficient throughput,low memory utilization,high power consumption and insufficient update performance.In terms of message classification,traditional TCAM cannot store range rules efficiently.Based on this problem,a multi-field TCAM coding technique based on hypercube is designed by taking advantage of the symmetry and regularity of hypercube.Through the comparison of simulation experiments,the encoding effect is 2 times more efficient than otherpopular TCAM encoding schemes,which greatly increases the space utilization of TCAM in message classification.
Cross-layer Matching Mechanism of Link Communication Rate for Heterogeneous Communication in Power System
XIAO Yong, JIN Xin, FENG Jun-hao
Computer Science. 2021, 48 (11A): 495-499.  doi:10.11896/jsjkx.200500113
Abstract PDF(2088KB) ( 566 )   
References | Related Articles | Metrics
Various local communication techniques are employed in the power system.The coexist of multiple protocols in power networks leads to various communication rates,resulting in difficulty of heterogeneous fusion.Aiming at this problem,a cross-layer communication rate matching mechanism between heterogeneous links is proposed.In the link layer,the link utilization rate is improved by means of the rate requirement based access control mechanism.Using the dynamic store-and-forward method,the communication rate is matched and forwarded among different network interfaces.Furthermore,in the network layer,data rate matching based on multipath transmission is proposed to achieve the maximum data rate matching and improve the communication efficiency.Simulation results have demonstrated the validity of the proposed rate matching mechanism based on cross layer design.In addition,compared with the RBAR protocol and ARF protocol,the proposed mechanism can be used to improve the network throughput,while reduce the delay and packet loss rate.
Information Security
Overview of Blockchain Technology
DAI Chuang-chuang, LUAN Hai-jing, YANG Xue-ying, GUO Xiao-bing, LU Zhong-hua, NIU Bei-fang
Computer Science. 2021, 48 (11A): 500-508.  doi:10.11896/jsjkx.201200163
Abstract PDF(2045KB) ( 2329 )   
References | Related Articles | Metrics
With the increasing popularity and development of virtual digital currencies such as bitcoin,blockchain technology has been widely concerned by researchers.Blockchain technology is a distributed data ledger which combines blocks in a chain structure according to the time sequence.It has the characteristics of decentralization,programmability,and traceability.It has been widely studied in the financial field.Facing the development of blockchain technology,this paper introduces the origin and overview of blockchain technology,discusses in detail the crucial technologies of blockchain,consisting of ring signature,zero know-ledge proof,digital signature and homomorphic encryption,and summarizes the characteristics and types of blockchain technology.This paper summarizes the application field of blockchain technology,focuses on its application principles and relevant cases in the application field,analyzes the current development status of blockchain application,and analyzes and forecasts the development direction of blockchain technology in the future.
Survey of Research Progress of Code Clone Detection
LE Qiao-yi, LIU Jian-xun, SUN Xiao-ping, ZHANG Xiang-ping
Computer Science. 2021, 48 (11A): 509-522.  doi:10.11896/jsjkx.210300310
Abstract PDF(2041KB) ( 1658 )   
References | Related Articles | Metrics
In a software system,if there are two or more similar pieces of code,it is called code clone.Studies have shown that code cloning exists in a large number of software systems and is increasing with the passage of time.In the era of big data,open source code has been a trend,so the proportion of code cloning will be higher and higher.The existing related work thinks that the code cloning in software system is harmful,which will lead to the decrease of system stability,the redundancy of code base and the spread of software defects.This paper analyzes the reasons,advantages and disadvantages of software cloning,and classifies the code cloning in software system.In order to improve the code quality,many code cloning detection methods have been proposed in the academic and industrial circles.According to the degree of information obtained from the code,these methods can be divided into five types based on text,lexis,syntax,semantics,and metric.This paper analyzes and evaluates these five kinds of detection methods,and finds that all kinds of methods have no absolute advantages,and different methods can be applied to diffe-rent fields according to different requirements.In the end,the paper summarizes the different application scenarios of the cloning detection technology,and forecasts the development direction of the code cloning detection methods and applications.
Study on Threat of Persistent Fault Attack
WANG Jian, CHEN Hua, KUANG Xiao-yun, YANG Yi-wei, HUANG Kai-tian
Computer Science. 2021, 48 (11A): 523-527.  doi:10.11896/jsjkx.210200138
Abstract PDF(2392KB) ( 620 )   
References | Related Articles | Metrics
Persistent Fault Attack(PFA) is a powerful attack which relies on persistent fault and statistical analysis,it can be applied in extracting secret key of block cipher implementation based on lookup tables.The greatest advantage of PFA is that it can recover the secret key with only one fault injection,meanwhile,it can be applied in countermeasures on fault attack like detection,mask and so on.However,these countermeasures still can make the attack more difficult,key recovery on implementation with countermeasures based on detection and infection need several times cipher text,this will hinder actual attack.Built-in health test for S-box will be a good countermeasure for PFA,the cipher device will stop working once there is a fault injection.PFA relies on the bijective characteristic of the S-box in block cipher,therefore,testing the bijection characteristic of S-box is an effective method to get a health test result for S-box.Just 255 XOR operations will give a reliable health test result for S-box,it costs much less than a normal test method like SHA3.Furthermore,non-algorithmic countermeasures like laser sensor should attractive some attention.
Security Cooperation of UAV Swarm Based on Blockchain
WANG Yu-chen, QI Wen-hui, XU Li-zhen
Computer Science. 2021, 48 (11A): 528-532.  doi:10.11896/jsjkx.201100199
Abstract PDF(1956KB) ( 1954 )   
References | Related Articles | Metrics
Unmanned aircraft vehicle swarm not only broadens the application fields of unmanned aircraft vehicles (UAVs),but also brings more complex security issues.Both the intelligence collected by UAVs and UAV itself are vulnerable to hacker attacks.Once hackers intercept communication information or hijack UAVs,it will have an adverse effect on the combat environment,and even cause serious consequences such as leaking state secrets and disrupting social order.This article introduces the security issues faced by UAV swarm and the principle of the consortium blockchain-Hyperledger Fabric.Then it designs a peer-to-peer network model named “UAV-Swarm Net” based on Hyperledger Fabric,using distributed ledgers and smart contracts to implement the safe cooperative combat of UAV swarm.The UAV-Swarm Net model has the characteristics of distribution,multiple copies,encrypted channel transmission and responsibility traceability,which can effectively guarantee the confidentiality,integrity and availability of data,enhance the robustness and scalability of UAV system.It provides a new solution to the information security problems faced by UAV swarm in the future UAV systemized combat environment.
Shared Digital Credits Management Mechanism of Enterprise Alliance Based on Blockchain
LING Fei, CHEN Shi-ping
Computer Science. 2021, 48 (11A): 533-539.  doi:10.11896/jsjkx.201200170
Abstract PDF(3514KB) ( 879 )   
References | Related Articles | Metrics
Traditional digital credits management mechanisms have decentralized,utilization restrictions,inconvenient exchange circulation and other problems,which limit the use of credits.Aiming at the existing problems of traditional digital credits mechanism,this paper puts forward the management mechanism of enterprise alliance sharing digital credits.Based on the decentralized mechanism of blockchain,secure identity authentication management,distributed database and smart contract automatic processing and other advantages,a shared digital credits management mechanism of enterprise alliance based on blockchain is designed.Combined with the application scenario of enterprise digital credits transaction,the performance of block chain network throughput and delay is tested and studied.The feasibility of the shared digital credits management mechanism of enterprise alliance is verified through simulation experiment,which provides reference for the enterprise-level application of block chain.
Network Anomaly Detection Based on Deep Learning
YANG Yue-lin, BI Zong-ze
Computer Science. 2021, 48 (11A): 540-546.  doi:10.11896/jsjkx.201200077
Abstract PDF(3342KB) ( 1370 )   
References | Related Articles | Metrics
This paper proposes a novel and general end-to-end convolutional transformer network for modeling the long-range spatial and temporal dependence on network anomaly detection.The core ingredient of the proposed model is the feature embedding module by just replacing the spatial convolutions with proposed global self-attention in the final three bottleneck blocks of a ResNet,and the multi-head convolutional self-attention layer in encoder and decoder,which learns the sequential dependence of network traffic data.Our model uses an encoder,built upon multi-head convolutional self-attention layers,to map the input sequence to a feature map sequence,and then another deep networks,incorporating multi-head convolutional self-attention layers,decode the target synthesized feature map from the feature maps sequence.We also present a class-rebalancing self-training framework to alleviate the long tail effect caused by the imbalance of data distribution through semi-supervised learning,which is motivated by the observation that existing SSL algorithms produce high precision pseudo-labels on minority classes.The algorithm iteratively retrains a baseline SSL model with a labeled set expanded by adding pseudo-labeled samples from an unlabeled set,where pseudo-labeled samples from minority classes are selected more frequently according to an estimated class distribution.In this paper,CIC-IDS-2017 datasets is used for experimental evaluation.The experiments shows that the accuracy of our model is higher than that of other deep learning models,which improves detection performance while reducing detection time,and has practical application value in the field of network traffic anomaly detection.
Block-chain Privacy Protection Scheme Based on Lightweight Homomorphic Encryption and Zero-knowledge Proof
WANG Rui-jin, TANG Yu-cheng, PEI Xi-kai, GUO Shang-tong, ZHANG Feng-li
Computer Science. 2021, 48 (11A): 547-551.  doi:10.11896/jsjkx.201200138
Abstract PDF(1981KB) ( 1265 )   
References | Related Articles | Metrics
In order to solve the problem of block-chain privacy protection and its efficiency,this paper proposes a privacy protection scheme based on lightweight homomorphic encryption and zero-knowledge proof.The scheme improves the homomorphic encryption algorithm to reduce the time of key generation and encryption,and introduces zero-knowledge proof to avoid invalid homomorphic operation.After the lightweight homomorphic encryption,the private data will be written into the block in the form of ciphertext,it is uploaded to the blockchain network by the node that gets the bookkeeping right.The scheme makes up for the lack of data disclosure in blockchain network and improves the efficiency.By analyzing the security of the scheme,it is proved that the scheme has the characteristics of unforgeability and privacy data security.Through the performance simulation experiment and theoretical deduction,it is proved that the low efficiency in the process of distributing,sharing and computing private data in ciphertext state has been improved,and it is more effective to protect the privacy of customers than the traditional DRM.
Analysis and Application of Secure Boot Algorithm Based on IOT Chip
ZONG Si-jie, QIN Tian, HE Long-bing
Computer Science. 2021, 48 (11A): 552-556.  doi:10.11896/jsjkx.210300237
Abstract PDF(1862KB) ( 846 )   
References | Related Articles | Metrics
RSA and ECC are currently standard public key encryption algorithm in IOT chips.By comparing the performance and security between RSA and ECC algorithms,ECC is found to be more suitable for IOT applications.A secure driver program is proposed as the solution of the secure startup of IOT chips.Experimental verification based on IOT chips demonstrates that ECC algorithm possesses several advantages including lower cost,higher performance,and better security.This paper provides a solution for the security development of the Internet of Things.
Natural Language Steganography Based on Neural Machine Translation
ZHOU Xiao-shi, ZHANG Zi-wei, WEN Juan
Computer Science. 2021, 48 (11A): 557-564.  doi:10.11896/jsjkx.210100015
Abstract PDF(2464KB) ( 857 )   
References | Related Articles | Metrics
Generation-based natural language steganography embeds secret information during text generation under the guidance of secret bitstream.The current generation-based steganographic methods are based on recurrent neural networks (RNN) or long short-term memory (LSTM),which can only generate short stego text because the semantic quality becomes worse as the length of the sentence increases.Moreover,there is hardly any semantic connection between sentences.To address this issue,this paper proposes a neural machine translation steganography algorithm,namely Seq2Seq-Stega,that can generate long text in which semantic relationship maintains well between words and sentences.An encoder-decoder model based on sequence-to-sequence (Seq2Seq) structure is used as our translation model.The source sentence can offer extra information and ensure the semantic relevance between the target stego sentences.In addition,according to the word probability distribution obtained by the model,we design a word selection strategy to form the candidate pool.An attention hyperparameter is introduced to balance the contribution of the source sentence and the target sentence.Experimental results show the hidden capacity and the text quality under different word selection thresholds and attention parameters.Comparative experiments with other three generation-based models show that Seq2Seq-Stega can maintain long-distance semantic connections and better resist steganalysis.
Camera Identity Recognition Method Fused with Multi-dimensional Identification Features
ZHU Rong-chen, LI Xin, WANG Han-xu, YE Han, CAO Zhi-wei, FAN Zhi-jie
Computer Science. 2021, 48 (11A): 565-569.  doi:10.11896/jsjkx.210100093
Abstract PDF(1731KB) ( 644 )   
References | Related Articles | Metrics
With the development of smart cities and public security big data,video surveillance networks have become essential infrastructure for urban governance.However,by replacing or tampering with surveillance cameras- the important front-end device,an attacker can access the internal network to achieve device hijacking,information theft,network paralysis,and threatening personal,social,and national security.A camera identity recognition method combining multi-dimensional identification features is proposed to identify illegal camera identities in advance.A camera identification system that integrates explicit,implicit,and dynamic identifiers is constructed by extracting the camera's static information and the dynamic flow information.An evaluation method of identifier contribution based on self-information and information entropy is proposed to select a concise and practical identity identifier.The extracted identifier feature vector can lay the foundation for future abnormal camera intrusion detection.Experimental results show that explicit identifiers have the most considerable amount of self-information and contribution but are easy to be forged;dynamic identifiers have the second-highest contribution,but the workload of traffic collection and processing is enormous;static identifiers have a low contribution but still have a specific role in identification.
Enterprise Risk Assessment Model Based on Principal Component Regression and HierarchicalBelief Rule Base
LIU Shan-shan, ZHU Hai-long, HAN Xiao-xia, MU Quan-qi, HE Wei
Computer Science. 2021, 48 (11A): 570-575.  doi:10.11896/jsjkx.201200038
Abstract PDF(2630KB) ( 598 )   
References | Related Articles | Metrics
As a new intelligent expert system with the characteristics of expert system and data-driven model,the belief rule base (BRB) plays an important role in risk assessment and health status assessment.BRB has the advantages of processing numerical data and linguistic qualitative knowledge from heterogeneous sources,which can help enterprises conduct effective risk assessments.However,in the actual enterprise risk evaluation system,there are many types of indicators and redundancy.Traditional BRB cannot select indicators and is easy to cause rule explosion,which leads to problems such as large calculation amount and low model accuracy.In response to the above problems,this paper proposes a principal component regression and hierarchical confidence rule base (Principal Component Regression,Hierarchical Belief Rule Base,PCR-HBRB) enterprise risk assessment model,which saves calculation time by screening effective indicators,and combining qualitative with quantitative information to analyze and evaluate,to obtain higher accuracy evaluation results.Firstly,the PCR method is used to screen out the main indicators that affect the states of the company,the hierarchical confidence rule base (HBRB) inference model of company status risk assessment is established based on the selected indicators,and the model is reasoned by the evidence reasoning (ER).Then,the projection covariance matrix adaptive evolution strategy (P-CMA-ES) is used to optimize the model.Finally,the effectiveness of the model is verified through a risk assessment case of a certain enterprise's financial situation.
Efficient Multi-keyword Retrieval Scheme Based on Attribute Encryption in Multi-cloud Environment
HE Heng, JIANG Jun-jun, FENG Ke, LI Peng, XU Fang-fang
Computer Science. 2021, 48 (11A): 576-584.  doi:10.11896/jsjkx.201000026
Abstract PDF(2604KB) ( 578 )   
References | Related Articles | Metrics
With the rapid development and wide application of cloud computing technology,data security issues in the cloud environment have become the focus of users' attention.To ensure data privacy,users encrypt the private data and upload it to the cloud server.Nevertheless,it is challenging to retrieve ciphertext containing specific information from massive encrypted data of multiple cloud servers.Traditional searchable encryption schemes cannot be directly applied to ciphertext data retrieval in the multi-cloud environment.The attribute-based encryption provides a new solution for ciphertext keyword retrieval.However,the existing related schemes have some problems,such as only supporting single or conjunctive keyword retrieval,inflexible access control policy,low retrieval efficiency,large calculation and storage overhead,and not applying to the multi-cloud environment effectively.Therefore,this paper proposed an efficient Multi-keyword Retrieval scheme based on Attribute encryption in the Multi-cloud environment (MRAM).MRAM is based on the high-performance ciphertext-policy attribute-based encryption algorithm,and realizes multi-keyword ciphertext retrieval and fine-grained access control.By introducing a retrieval server,MRAM effectively supports efficient and accurate ciphertext retrieval in multi-cloud environment.Security analysis shows that MRAM can achieve important security features such as security index confidentiality,trapdoor confidentiality,and resistance to collusion attacks.The performance evaluation verifies that MRAM has lower computational overhead in the secure index generation,trapdoor generation,and retrieval stages compared with existing solutions,and the storage overhead of the secure index and trapdoor is also smaller.
Optimization Study of Sketch Algorithm Based on AVX Instruction Set
TAN Ling-ling, YANG Fei, YI Jun-kai
Computer Science. 2021, 48 (11A): 585-587.  doi:10.11896/jsjkx.210100205
Abstract PDF(2004KB) ( 644 )   
References | Related Articles | Metrics
Network traffic detection is the most basic and important part of network measurement.The detection method based on Sketch can count data flow of network,which can distinguish elephant flow,detect and locate abnormal traffic in abnormal detection of network traffic.Higher requirement of the memory resources is needed for multiple hash functions used in the implementation of Sketch.To address the restriction problems raised by multiple hash functions applied in the realization of Sketch,a me-thod of improving the performance of Sketch algorithm using AVX instruction set is put forward,and also the effects on CPU instruction consumption and algorithm computing efficiency are studied,which had high efficiency and high maturity.Firstly,the elements of data stream are described and stored in the form of a vector,utilizing AVX instructions to realize structures and opera-tions of vectors.Secondly,multiple hash operations are transformed into a vector operation,which reduces the CPU resources consumption of the Sketch algorithm,improves the comprehensive performance of multiple hash functions,and made it possible to improve the optimized performance of Sketch algorithm.Finally,AVX instruction set is used to optimize the key functions in the Count-Min Sketch algorithm program,and the optimized code is tested and analyzed.The experimental results show that the AVX optimized version takes 25% time of the original version in the operation of hash function.When the data length is short,the instructions consumed by multiple hash functions are relatively small in the whole Sketch,and the AVX optimized version consumes about 70% time of the original version.As the length of the data increases,instructions consumed by multiple hash functions take up a larger proportion in the whole Sketch algorithm,and the time consumed by the AVX optimized version can be reduced to 40% of the original version.The simulation results demonstrate the effectiveness of the proposed algorithm in improving the measurement efficiency of network data flow.
Application of Express Information Encryption Based on AES and QR
GU Shuang-jia, LIU Wan-ping, HUANG Dong
Computer Science. 2021, 48 (11A): 588-591.  doi:10.11896/jsjkx.210100024
Abstract PDF(2543KB) ( 690 )   
References | Related Articles | Metrics
With the rapid development of e-commerce,the express industry is particularly hot.A large number of users habitually throw away the unprocessed express box after taking the express,at this time,a large number of users' plaintext privacy information on the express bill will face the risk of leakage.AES algorithm is a symmetric cipher algorithm,which has the characteristics of simple algorithm,fast encryption and decryption,high security and so on.It is an ideal algorithm for encrypting express industry users' plaintext information.QR code (fast response two-dimensional code) has the advantages of high security,fast reading speed,printing in small space,and reading from any direction.In order to ensure the efficiency of transportation,express delivery is based on the user's plaintext information.After the express delivery arrives at the destination,AES algorithm is used to encrypt the user's private plaintext information into ciphertext and generate the corresponding ciphertext two-dimensional code,which is printed and pasted on the express box.Python is used to realize the generation and recognition system of ciphertext two-dimensional code data,which not only protects the information security of users,but also reduces the phenomenon of pretending to get express delivery,which has high practicability.
Network Intrusion Detection System Based on Multi-model Ensemble
MA Lin, WANG Yun-xiao, ZHAO Li-na, HAN Xing-wang, NI Jin-chao, ZHANG Jie
Computer Science. 2021, 48 (11A): 592-596.  doi:10.11896/jsjkx.201100170
Abstract PDF(2266KB) ( 652 )   
References | Related Articles | Metrics
The network intrusion detection system (NIDS) is widely used in the construction of network security.It can effectively identify the potential behaviors that endanger network security.In order to obtain more accurate and efficient network intrusion detection results,a network intrusion detection system based on multi-model ensemble is proposed.The system integrates Linear Support Vector Machines (Linear SVM),Residual Networks (NETS) and Temporal Convolutional Network (TCN) by using Bagging algorithm to detect the network intrusion.Intrusion detection data in experiments are 99809 web log data and AWIDof work equipment in State Grid Shandong Electric Power Companyas its public data sets.This system is compared with the single use Linear SVM,ResNets,TCN this three model.The experimental results show that by using multi-model ensemble algorithm,integrating the advantages of each model,the overall accuracy of this system reaches up to 99.24% and is 7.95% more than TCN.In addition,the system not only has a very high accuracy rate,the alarm rate is also as low as 0.07%,which is consistent with the requirements of network security protection system,and successfully realizes more accurate and efficient network intrusion detection.
Interdiscipline & Application
Review on Intelligent Diagnosis of Spine Disease Based on Machine Learning
LIU Tong-tong, YANG Huan, XI Yong-ming, GUO Jian-wei, PAN Zhen-kuan, HUANG Bao-xiang
Computer Science. 2021, 48 (11A): 597-607.  doi:10.11896/jsjkx.201100006
Abstract PDF(1949KB) ( 1274 )   
References | Related Articles | Metrics
Spine diseases are prevalent in modern society.The diagnosis and treatment mainly depend on doctors' professional knowledge and clinical experience.More and more patients and conventional treatments resulted in heavy overload and inefficient diagnosis.Machine learning algorithms can automatically extract useful information from datasets and images,assisting doctors to locate the lesion and carry out the accurate treatment.This paper focuses on the applications of machine learning in the field of spine disease and summarizes the relevant research from aspects of datasets,feature selection,model,evaluation metrics,and so on.Firstly,in terms of machine learning algorithms,the utility of some typical algorithms in disease diagnosis and treatment is described.Moreover,in terms of the actual applications of disease diagnosis and treatment(risk factor analysis and disease prediction,disease recognition and classification,feature extraction of spine image and image segmentation),the performances of several important models are compared in some specific experiments.Accuracy,specificity,sensitivity,AUC,and other evaluation indexes are involved.Finally,the major limitations and corresponding issues in current applications are summarized.
Meta-learning Algorithm Based on Trend Promote Price Tracing Online Portfolio Strategy
DAI Hong-liang, LIANG Chu-xin
Computer Science. 2021, 48 (11A): 608-615.  doi:10.11896/jsjkx.201100068
Abstract PDF(2551KB) ( 1166 )   
References | Related Articles | Metrics
With the rapid development of China's economy and the continuous increase of income available for distribution,people's investment demand has become more intense,and how to efficiently and rationally carry out investment portfolio has become a hot issue for investors.In response to the problem that online portfolio strategy predict stock price singly and is difficult to determine its accurate investment proportion,we propose a meta-learning algorithm based on Trend Promote Price Tracing (TPPT).Firstly,considering the influence of stock price anomalies,a three-state price prediction method which is based on the equal-weight slope value of the historical window period,is proposed to track the price changes.Secondly,the Error Back Propagation (BP) algorithm based on gradient projection is added to solve the investment ratio.Thus,TPPT strategy maximizes the cumulative wealth by feeding back the increasing capacity of assets to the investment ratio.Finally,the empirical analysis of five typical data shows that TPPT strategy has a great advantage in balancing risk and return,it is shown that TPPT strategy is a robust and effective online portfolio strategy.
U-net for Pavement Crack Detection
Computer Science. 2021, 48 (11A): 616-619.  doi:10.11896/jsjkx.201200059
Abstract PDF(2392KB) ( 1105 )   
References | Related Articles | Metrics
Road is one of the most crucial ways for transportation.Crack on road will cause great danger to transportation if you leave it unchecked,so it is important to detect crack precisely in road maintenance.Road cracks are usually discontinuous and low-contrast which is difficult to detect using traditional methods of image processing.In this paper,we utilize U-net for road crack detection which is an end-to-end with encoder-decoder structure efficient deep learning network on dataset Crack500,while traditional methods are time-consuming and labor-consuming.U-net is appropriate for road crack detection because of its ability to catch fine details in image.Experiment results demonstrate that U-net outperforms other detect methods.Furthermore,we discuss the difference when modifying the number of conv-blocks in U-net.Experiment results show that it achieves best performance when the number of conv-blocks set to be 7.
Image Quality Assessment for Low-dose-CT Images of COVID-19 Based on DenseNet and Mixed Domain Attention
SUN Rong-rong, SHAN Fei, YE Wen
Computer Science. 2021, 48 (11A): 620-624.  doi:10.11896/jsjkx.201200252
Abstract PDF(2562KB) ( 582 )   
References | Related Articles | Metrics
It is important to study the image quality assessment algorithm of low-dose-CT images for COVID-19.However,with the increase of the number of network layers,the gradient will disappear for the method based on deep learning.To solve this problem,this paper proposes DenseNet algorithm based on mixed domain attention.DenseNet solves the problem of gradient vanishing while reducing parameters through feature reuse and tight connection of network.Based on the attention mechanism of human vision,it adopts the combination of bottom-up and top-down structure to realize spatial attention.Based on the multi-channel characteristics of human vision,this paper ignores the information in the channel domain for spatial attention,studies and introduces the mixed domain attention to DenseNet.Spearman rank order correlation coefficient and Pearson linear correlation coefficient are used to measure the consistency between objective assessment method and subjective assessment method.The experimental results show that this method can simulate the human visual characteristics better,and evaluate the quality of low-dose-CT more accurately.The evaluation results are in good agreement with the subjective feelings of human vision.
Research on Application of Improved GAN Network in Generating Short Video
YU Xiao-ming, HUANG Hua
Computer Science. 2021, 48 (11A): 625-629.  doi:10.11896/jsjkx.210300114
Abstract PDF(2654KB) ( 762 )   
References | Related Articles | Metrics
In the study of the dynamic image generated by GAN,there are many problems,such as inconsistent colors of some objects and unnatural details of generated images.In order to solve the problem of unsatisfactory video generation,the main schemes adopted are to improve the generator and discriminator of GAN network respectively,which are shown in two aspects.On the one hand,the foreground and background of the videos are modeled separately in the generator and the Multi Spatial-Adaptive Normalization (M-Spade) algorithm is used.The other aspect is the use of dual video discriminator (DVD-GAN) on discriminator selection,which trained on Kinetics 600 dataset.The experimental results are compared with F-VID2VID,WC-VID2VID and other generation methods.The results show that the method of combining the two methods has a great effect on the problem of color inconsistency before and after the short video and the details processing,and the generated images are relatively clearer.
Diagnostic Prediction Based on Node2vec and Knowledge Attention Mechanisms
LI Hang, LI Wei-hua, CHEN Wei, YANG Xian-ming, ZENG Cheng
Computer Science. 2021, 48 (11A): 630-637.  doi:10.11896/jsjkx.210300070
Abstract PDF(2223KB) ( 777 )   
References | Related Articles | Metrics
Diagnostic prediction predicts the future diagnosis of patients from their historical health states,and it is the core task of personalized medical decisions.Electronic health record(EHR) documents patients' time-varying health conditions and clinical care,and also provides a wealth of longitudinal clinical data for diagnostic prediction.However,the existing diagnostic prediction models based on EHR can not completely learn the hidden disease progression patterns.Moreover,the performance of fine-grained diagnostic prediction greatly depends on more informative sequence features.In order to improve the performance,we propose adiagnostic prediction model,called Node2vec and knowledge attention model (NKAM).Specifically,based on Node2vec,the model captures the potential medical knowledge from the global structure of medical ontology.It also maps categories into low-dimensional vectors and encodes the medical knowledge of patients' health state into category embedding vectors.The diagnosis code embedding vectorsare used to enrich the patients' fine-grained health state representation.Then,the long-term dependencies and disease progression patterns can be extracted from the patient's historical health states using a knowledge attention mechanism combined with the Gated Recurrent Unit(GRU).Experimental results on real-world dataset show that NKAM significantly improves the prediction performance compared with state-of-the-art methods.Furthermore,the experiments reveal that Node2vec can capture more informative medical concept embedding from medical ontology,and the knowledge-based attention mechanism helps to the effective integration of external knowledge and electronic health records.
Prediction Model of Bubble Dissolution Time in Selective Laser Sintering Based on IPSO-WRF
ZHANG Tian-rui, WEI Ming-qi, GAO Xiu-xiu
Computer Science. 2021, 48 (11A): 638-643.  doi:10.11896/jsjkx.210300080
Abstract PDF(3056KB) ( 531 )   
References | Related Articles | Metrics
In order to solve the problem of mass defects caused by bubbles during selective laser sintering (SLS),a weighted random forest (WRF) prediction method based on improved particle swarm optimization algorithm (IPSO) was proposed to predict the dissolution time of bubbles effectively.In this method,two key parameters of WRF,the number of split attributes and the number of decision trees,were optimized by IPSO algorithm to construct the prediction model of IPSO-WRF.Numerical examples show that compared with the prediction results of PSO-RF and PSO-KELM prediction models,IPSO-WRF prediction model based on the same training samples and test samples can get the output results of bubble dissolution time with smaller error and closer to the actual value.MAE,MAPE and RMSE indexes show that IPSO-WRF prediction model has higher nonlinear fitting ability and prediction accuracy than PSO-RF model and PSO-KELM model.Finally,the most significant input parameters affecting the bubble dissolution time are determined by sensitivity analysis,which provides a theoretical basis for the development of SLS technology.
Three-dimensional Reconstruction of Cone Meteorological Data Based on Improved MarchingTetrahedra Algorithm
MA Jun-cheng, JIANG Mu-rong, FANG Su-qin
Computer Science. 2021, 48 (11A): 644-647.  doi:10.11896/jsjkx.210200025
Abstract PDF(3257KB) ( 768 )   
References | Related Articles | Metrics
In the field of meteorology,the meteorological data detected by Doppler radar is stored in the form of spatial polar coordinates,and the detected meteorological targets have the characteristics of uneven distribution,scattered regions,and irregular shapes.In order to meet the needs of 3D reconstruction of meteorological targets,the Marching Tetrahedra 3D reconstruction algorithm is improved according to the characteristics of radar data.First,the Barnes interpolation method and the interpolation method of the Fourier spectrum analysis principle are used in the vertical direction and radius of the radar cone data.This algorithm encrypts the echo intensity value between the two directions,divides the new hexahedron formed by the encrypted echo polar coordinate data into basic tetrahedral units,and uses linear interpolation to obtain the specific position of each vertex,and combines the multi-level surface when drawing.The rendering technology renders 3D images.The algorithm avoids the reconstruction of areas with high elevation angles and long distances without echo data.Experiments show that the improved algorithm can not only achieve better and faster three-dimensional reconstruction,but also observe and analyze the internal details of the cloud layer,which provides a certain reference basis for accurate weather forecasting.
Fault Detection for Arc Magnet Based on Convolutional Neural Network and Acoustic VibrationImage
LIU Xin, HUANG Qin-yuan, LI Qiang, RAN Mao-xia, ZHOU Ying, YANG Tian
Computer Science. 2021, 48 (11A): 648-654.  doi:10.11896/jsjkx.210100161
Abstract PDF(4584KB) ( 650 )   
References | Related Articles | Metrics
As a key component in permanent magnet motor,the product quality of arc magnet is susceptible to degradation due to internal defects.However,traditional acoustic vibration detection methods have revealed some inefficiencies in the face of fast and accurate inspection requirements,so it is of great practical importance to develop an efficient and intelligent detection method for internal defects in arc magnets.This paper combines the advantages of deep learning and proposes a convolutional neural network-based acoustic vibration detection method for internal defects of arc magnets.In this method,the one-dimensional acoustic vibration signal of the arc magnets is firstly converted into the two-dimensional acoustic vibration image,and then fed into a convolutional neural network designed for the signal characteristics for learning and training,to complete the autonomous learning from the acoustic vibration image and extract the features that can distinguish the presence or absence of internal defects.Finally,the corresponding features are identified by softmax.The experimental results of four types of arc magnet samples show that the proposed method can achieve 99.38% accuracy of internal defect detection of arc magnets,the detection time of a single arc magnet is less than 0.031 s and has a high robustness of the model.
Research on Behavior of Travel Mode Choice Under Duration of Public Health Emergencies Basedon SEM Model
LUO Chen, HU Ming, XIAO Huan-quan, ZHONG Lin-feng
Computer Science. 2021, 48 (11A): 655-658.  doi:10.11896/jsjkx.201200004
Abstract PDF(1702KB) ( 738 )   
References | Related Articles | Metrics
In order to analyze the decision-making mechanism of the choice of residents' travel mode during the duration of public emergencies,the structural equation model is introduced to construct the SEM model of residents' travel mode selection,based on the 623 effective surveys collected through the network,during the duration of the cronavirus disease 2019 (COVID-19) in march 2020.The questionnaire investigates the main influencing factors of residents' travel choices during the duration of public emergencies.The results show that family income per capita,whether there is a private car,prevention and control measures,and the degree of prevention and control measures have absolute impact on residents' travel modes.Reducing stopover places in the process of high-speed rail passenger transport,and keeping low ticket price in air passenger transport,can improve the probability of residents choosing corresponding travel modes.It can be seen that the behavior choice of residents during the duration of public health emergencies is quite different from that under normal conditions.
Subprocesses Discovery Based on Structure and Activity Semantics
SUN Shan-wu, WANG Nan
Computer Science. 2021, 48 (11A): 659-665.  doi:10.11896/jsjkx.210100089
Abstract PDF(1903KB) ( 456 )   
References | Related Articles | Metrics
More companies document their business operations in the form of process models,and require descriptions of one process on various levels of detail.Given a detailed process model,business process model abstraction (BPMA) delivers abstract representations for the same process to achieve particular abstraction goals.A prominent BPMA use case is a construction of a process “quick view” for rapidly comprehending a complex process.A key problem in this abstraction scenario is the transition from detailed activities in the initial model to coarse-grained activities.Many researchers have studied the methods of process abstraction most of which are based on structure.The structure-based abstraction derives the set of activities to be abstracted from the original model according only to control flow relations,but not considering the domain semantics,resulting in many process fragments (candidate subprocess) with incomplete business logicality.This paper bases on the structure of process and activity semantics to extend every canonical component of the process structure tree from the bottom up.This paper assesses the similarity between the canonical component and all its adjacent nodes to find out the group of activities which is the most similar to the canonical component.And then it aggregates the generated group of activities to derive the process segments as candidate subprocesses.This paper uses real-running process cases to compare structure-based abstraction method which takes every canonical component as a candidate subprocess with the proposed method.The results show that the number of the subprocesses with uncompleted business meaning is greatly reduced and the generated groups of activities are more similar to manually designed subprocesses.
On Weakly Process of Bounded Petri Net and Weakly Occurrent Nets
LIU Ping
Computer Science. 2021, 48 (11A): 666-668.  doi:10.11896/jsjkx.210100127
Abstract PDF(1586KB) ( 477 )   
References | Related Articles | Metrics
The full process (N,φ) of bounded Petri nets Σ provides a powerful tool to study the reachable remarked of bounded Petri nets Σ,mapping the S cut of occurrence net N to the reachable remarked of Σ by using occurrent net N and net map φ.Since there is no more than one transition in the back-set of any space in the occurrence network,when there are more than one transition in the back-set of some space of the Petri nets Σ,the phenomenon of multiple repeats will appear.It makes the calculation too complication.In this paper,the weakly process of bounded Petri net and weakly occurrent nets are introduced to replace full process of Petri nets Σ.The main results hold in full process are proved hold in weak processes of bounded Petri nets.Thus weak processes are a meaningful extension of full processes.Because the weak occurrence network cancels the restriction on the number of elements in the backset of the space in the weak process,the repetition phenomenon caused by the above reasons in the full process is eliminated.So that the efficiency of calculation are improved.The example given in this paper shows the simplicity of weak occurrence networks in calculation.
Study on Axiomatic Truth Degree in First-order Logic
HAO Jiao, HUI Xiao-jing, MA Shuo, JIN Ming-hui
Computer Science. 2021, 48 (11A): 669-671.  doi:10.11896/jsjkx.210200012
Abstract PDF(1714KB) ( 533 )   
References | Related Articles | Metrics
First-order logic,as one of the standard formal logics of axiomatics systems,contains the research contents such as degree reasoning,which is a hot and difficult point.Based on the semantic theory of first-order logic calculus,this paper studies the axiomatic truth degree of first-order logic by using the satisfiability and completeness theorem.Firstly,the definition of the satisfiability of union and intersection operation is given.Secondly,the relationship between two special formulas and logical effective formulas and theorems is explained.Finally,the prenex normal form equivalent to formulas is obtained.The above results will prepare for the research of predicate logic degree.
Simulation Optimization and Testing Based on Gazebo of MPI Distributed Parallelism
JIANG Hua-nan, ZHANG Shuai, LIN Yu-fei, LI Hao
Computer Science. 2021, 48 (11A): 672-677.  doi:10.11896/jsjkx.210100109
Abstract PDF(2317KB) ( 972 )   
References | Related Articles | Metrics
Gazebo,as a general robot simulation platform,can simulate robot behavior accurately in the complex environment of indoor or outdoor,and support multi-robot collaborative simulation on single computer node.But when the simulation task contains hundreds of robots,it is usually found that the RTF (Gazebo simulation real-time performance) will reduce two orders of magnitude,some errors even appear in the simulation.The simulation performance will become the critical limiting factor.In order to realize high-performance simulation,the across node simulation platform based on MPI and ROS+Gazebo is explored.The core process is to divide the simulation tasks in parallel,which can be divided by number or region.The divided sub tasks are deployed to the Gazebo of each computing node for simulation.Finally,the MPI process communication between the Gazebo ensures the synchronization and consistency of the simulation,so as to realize the collaborative simulation of robots distributed on different computing nodes.At the same time,two types of cases including homogeneity and heterogeneity about fixed wing and quadrotor are writed,which are realized by reading the world configuration file and roslaunch file through the script program.The user-friendly starting mode similar to ROS was designed,and the single-node and cross-node performance tests are carried out to verify the advantage of distributed parallelism simulation.
Application of Chinese Cryptographic Algorithm in RPKI
LENG Feng, ZHANG Ming-kai, YAN Zhi-wei, ZHANG Cui-ling, ZENG Yu
Computer Science. 2021, 48 (11A): 678-681.  doi:10.11896/jsjkx.210100030
Abstract PDF(2206KB) ( 743 )   
References | Related Articles | Metrics
The security of routing systems attracts extensive attention worldwide with increasing inter-domain routing hijacking incidents in recent years.As a routing security verification system,the RPKI system can greatly reduce the risk of routing hijacking by working with existing routing broadcast strategies.The signature algorithm is limited to the RSA asymmetric encryption algorithm,and the hash algorithm is limited to the SHA-256 algorithm.With the upgrading of cryptographic algorithms,it is reasonable to expected that the RPKI system will gradually accept more algorithms to meet security and performance requirements.This article introduces the SM2 and SM3 algorithms,also known as Chinese commercial cryptographic algorithms,into RPKI system,and establishes a complete set of cryptographic algorithm testing environment from multi-dimensional aspect to compare Chinese commercial cryptographic performance with standard RFC defined algorithms.After performance evaluation and comparison,we discuss the algorithm feasibility,optimization and improvement methods in large-scale deployment environments,and the prospect of the future development of the existing crypto system in RPKI system.
Complex Algorithm Design and Maintenance Based on Thinking Map
ZHU Ping
Computer Science. 2021, 48 (11A): 682-687.  doi:10.11896/jsjkx.210100065
Abstract PDF(2291KB) ( 505 )   
References | Related Articles | Metrics
Because the software requirement specification focuses on the user requirements,it is not easy to describe and modify the internal logic of the complex algorithm.The flow chart is inclined to the bottom of the program design,it is convenient for the implementation of automatic program design,but lacks of high-level semantics,and has no mechanism to modify,update and backup,the logic complexity of modification is still very large,the cost of flow chart design and maintenance is also very high,and it is still not convenient to describe the uncertain program logic.To solve the practical problems in the designing and maintaining complex algorithms,this paper proposes a light-weight method to describe the internal logic of complex algorithm based on thinking map,which adapts to the characteristics of uncertain internal workflow and long-term R&D process.This paper formally defines the logical model of thinking map firstly,then discusses the object-oriented implementation of thinking map,as well as the program simulation for the inheritance and polymorphic scene of the objective world.And it takes the semantic recognition of text data element variables as an example,specifically describes the stage and process of designing and maintaining complex algorithms by thinking map.Finally,it summarizes the whole paper and puts forward some suggestions,the next research task of thinking map is also proposed
Research on Intelligent Control Technology of Accurate Cost for Unit Confirmation in All Links of Power Transmission and Transformation Project Based on Edge Computing
LUAN Ling, PAN Lian-wu, YAN Lei, WU Xiao-lin
Computer Science. 2021, 48 (11A): 688-692.  doi:10.11896/jsjkx.201100200
Abstract PDF(2923KB) ( 493 )   
References | Related Articles | Metrics
In order to further improve the refinement level of the cost management of power transmission and transformation projects,the precise cost control technology confirmed by the whole link unit of power transmission and transformation projects is studied in-depth.First of all,the existing problems such as the lack of accurate measurement method,lack of mature control technology and the difficulty in coordinating the caliber of each link unit are analyzed in detail.Then,the intelligent control model of accurate cost of the whole link unit confirmation of power transmission and transformation project based on edge computing is constructed.The model is optimized by immune particle swarm optimization algorithm based on hybrid strategy.The cost calculation model of each link of power transmission and transformation project built by edge computing greatly shortens the calculation delay time and reduces the cost of data redundancy.Finally,immune particle swarm optimization algorithm is adopted to optimize the model to get rid of the disadvantage that traditional optimization algorithm is easy to fall into local optimal.Immune particle swarm optimization makes model data processing more efficient and accurate.The algorithm further realizes the advantage of high reliability of edge computing collaboration,and realizes the high precision cost control system of each unit in the whole link.
Portfolio Optimization System Based on Multiple Trend Indices with Time Picking of Inducing Peak Prices
CHEN Jing-bang, PAN Jun-zhe, SHEN Hao-lang, GU Pei andHU Ming-tao
Computer Science. 2021, 48 (11A): 693-698.  doi:10.11896/jsjkx.210300215
Abstract PDF(2644KB) ( 817 )   
References | Related Articles | Metrics
Trend representation index is an important topic in the field of portfolio optimization.However,most of the portfolio optimization systems based on trend representation only consider one index,and the effect of the system considering only one index is often quite different on different data sets,so we use multiple trend indices in our system.The portfolio optimization system proposed in this paper uses a series of radial basis functions corresponding to three trend representation indices (simple mo-ving average line,exponential moving average line and low-lag trendline) respectively.This system uses the above three indices and adds the peak price index according to the relationship between the closed price and the short-term average price.In this system,the series of radial basis functions will select the best trend expression index (adaptive selection) according to the recent investment situation.Then,the system will make investment according to the solution set of the convex optimization problem which aims at maximizing the wealth of the next period.Finally,the system and five common portfolio optimization systems are compared on two data sets,two of which are chosen to be compared in more detailed on four data sets,and we conclude that our system is better than other systems.
Implementation and Optimization of Sunway1621 General Matrix Multiplication Algorithm
LI Shuang, ZHAO Rong-cai, WANG Lei
Computer Science. 2021, 48 (11A): 699-704.  doi:10.11896/jsjkx.201200150
Abstract PDF(3233KB) ( 931 )   
References | Related Articles | Metrics
As the most basic library in high performance computing (HPC),BLAS plays an important role in scientific computation,AI and other applications.The GEMM-based level 3 BLAS is the core of the performance of the entire BLAS.At present,there is no high-performance BLAS library that can give full play to the advantages of Sunway1621.Aiming at the above problems,we realize the transplantation and optimization of GotoBLAS on Sunway1621.This paper presents an algorithm for core code optimization using SIMD vectorization,and performs optimization techniques such as data regrouping,blocking,register allocation,and vectorization instruction optimization.The optimal data block selection scheme using vectorization and cache-based optimization for SGEMM and DGEMM in Micro-Kernel is compared respectively.Our optimizations achieve an average speedup of 52.09X and 32.75X on single precision and double precision compared to GotoBLAS.
Design of Intrusion Detection System Based on Sampling Ensemble Algorithm
HUAN Wen-ming, LIN Hai-tao
Computer Science. 2021, 48 (11A): 705-712.  doi:10.11896/jsjkx.201100101
Abstract PDF(4976KB) ( 590 )   
References | Related Articles | Metrics
As the second line of defense after firewalls,intrusion detection systems have been widely used in the field of network security.Machine learning-based intrusion detection systems have attracted more and more interest due to their superior detection performance.In order to improve the detection performance of the intrusion detection system in multiple types of imbalanced data,this paper proposes an intrusion detection system based on the optimal sampling ensemble algorithm(OSEC).OSEC first converts the multi-category detection problem into multiple binary classification problems according to the “one-to-all” principle,and then selects the optimal sampling ensemble algorithm according to the AUC value in each binary classification problem to alleviate the data imbalance problem.Finally,the category judgment module designed in this article judges the specific category of the sample to be tested.We perform simulation verification on the NSL-KDD data set,and find that compared with the traditional method,the F1 score of this system on R2L and U2R has increased by 0.595 and 0.185 respectively;compared with the latest intrusion detection system,the method in this paper improves the overall detection accuracy by 1.4%.
Research on Personnel File Management System Based on Blockchain
WANG Hui, CHEN Bo, LIU Yu-xiang
Computer Science. 2021, 48 (11A): 713-718.  doi:10.11896/jsjkx.210300051
Abstract PDF(2661KB) ( 986 )   
References | Related Articles | Metrics
According to the current situation of personnel file management in China,this paper puts forward a scheme of personnel file management system based on blockchain,and introduces the improved PBFT consensus algorithm and system scheme design from the system framework.The system uses the improved PBFT consensus algorithm to store the file data safely and effectively in the personnel file blockchain system,which ensures the traceability and effectiveness of the file data and uses technologies such as intelligent contract and inter planetary file system (IPFS) to realize the local backup and transfer sharing of personnel files,so as to prevent the malicious damage of the file data by the untrusted third party and ensure the security of the system.Experimental data show that the improved PBFT consensus algorithm has obvious advantages in output performance,consensus speed and security compared with other mature consensus algorithms,provides higher security and throughput.Experimental analysis shows that the personnel file management system based on blockchain is expected to change the disadvantages of traditional personnel file management to meet the growing demand of personnel file data protection and sharing.