Computer Science
Started in January,1974（Monthly）
Supervised and Sponsored by Chongqing Southwest Information Co., Ltd.
ISSN 1002-137X
CN 50-1075/TP
CODEN JKIEBK
Authors
Reviewers
Editors
Current Issue
Volume 49 Issue 11A, 16 November 2022
  
Artificial Intelligence
Cooperation and Confrontation in Crowd Intelligence
ZHU Di-di, WU Chao
Computer Science. 2022, 49 (11A): 210900249-7.  doi:10.11896/jsjkx.210900249
Abstract PDF(2277KB) ( 1725 )   
References | Related Articles | Metrics
Crowd intelligence has rich connotations and denotations.Its algorithms include both the early algorithms based on the characteristics of biological groups(particle swarm optimization,ant colony algorithm,etc.) and the later large-scale crowd algorithms based on network interconnection(multi-agent system,crowd intelligence perception,federated learning,etc.).The core idea of these crowd intelligence algorithms is cooperation or confrontation.Collaboration can combine the limited intelligence of individuals into the powerful intelligence of the group.However,collaboration itself has certain limitations,which may lead to the over-dependence between individuals and the unfairness of the system.Confrontation can overcome this limitation,and its basic idea is that individuals seek their maximum interests through the game.Therefore,cooperation and confrontation are indispensable.It is the inevitable development trend of a crowd intelligence to promote cooperation with confrontation,and to build a crowd intelligence ecology in which cooperation and confrontation coexist.This paper mainly focuses on the cooperation and confrontation methods of crowd intelligence algorithms,expounds on the classical crowd intelligence algorithms,and prospects the next development direction of emerging crowd intelligence algorithms.
Survey of Deep Learning Technologies for Financial Technology
ZHOU Fan, CHEN Xiao-die, ZHONG Ting, WU Jin
Computer Science. 2022, 49 (11A): 210900016-17.  doi:10.11896/jsjkx.210900016
Abstract PDF(2406KB) ( 1380 )   
References | Related Articles | Metrics
In recent years,deep learning techniques have been widely applied in addressing various problems in financial technology(Fintech) and have attracted increasing attention from both academia and business.Researchers utilize deep learning techniques for mining and analyzing financial data while finding the economic patterns behind tremendous data.Deep learning outperforms traditional statistical machine learning models in a range of crucial financial applications,including market movement prediction,trading strategy improvement,financial text processing,etc.To facilitate the development of Fintech and the deployment of new deep learning techniques,this paper provides a comprehensive survey of the deep learning-based Fintech studies published in recent years.Our survey focuses on the most recent advances in Fintech and provides a roadmap of financial problems as well as corresponding solutions.To this end,we investigate the widely used methodologies in finance data mining and summarize the popular deep models in Fintech data learning.Besides,we propose a taxonomy that categorizes existing Fintech research into ten well-studied applications in the literature.Subsequently,we systematically review the state-of-the-art deep learning methods and provide insights on the improvement for future endeavors.Finally,the pros and cons of existing research are summarized,followed by outlining the trend,open challenges,and opportunities in the Fintech research community.
Analysis of Technology Trends Based on Deep Learning and Text Measurement
WEI Ru-ming, CHEN Ruo-yu, LI Han, LIU Xu-hong
Computer Science. 2022, 49 (11A): 211100119-6.  doi:10.11896/jsjkx.211100119
Abstract PDF(3429KB) ( 1042 )   
References | Related Articles | Metrics
Traditionally,technical trend analysis tasks need to be done by experienced analysts,involving a lot of literature review and data analysis work,which is time-consuming and labor-intensive.Facing the above problems,this paper proposes a technology trend analysis model based on deep learning and text measurement,and a domain specific named entity recognition(NER) algorithm based on the BERT_BiLSTM_CRF model is designed with optimized masking mechanism.Taking news and literatures texts in the field of integrated circuit as data set,a comparative study between BiLSTM_CRF,BERT_BiGRU_CRF and the optimized BERT_BiLSTM_CRF* model proposed in this paper is carried out.The performance of NER is compared and analyzed.Compared with other algorithms,the proposed algorithm reaches 88.6%(measured by F1 value),laying the foundation for technical trend analysis.Based on the characteristics of knowledge graphs that relationships can be naturally expressed,an innovative method that combines knowledge graphs with text measurement technology is proposed,and the results of technical trend analysis are visualized from various perspectives,and ultimately assist analysts to carry out intelligent analysis of technical trends.
Survey of Research on Extended Models of Pre-trained Language Models
Abudukelimu ABULIZI, ZHANG Yu-ning, Alimujiang YASEN, GUO Wen-qiang, Abudukelimu HALIDANMU
Computer Science. 2022, 49 (11A): 210800125-12.  doi:10.11896/jsjkx.210800125
Abstract PDF(2450KB) ( 1170 )   
References | Related Articles | Metrics
In recent years,the proposal of Transformer neural network has greatly promoted the development of pre-training technology.At present,pre-training models based on deep learning have become a research hotspot in the field of natural language processing.Since the end of 2018,BERT has achieved optimal results in multiple natural language processing tasks.A series of improved pre-training models based on BERT have been proposed one after another,and pre-training model extension models designed for various scenarios have also appeared.The expansion of pre-training models from single-language to tasks such as cross-language,multi-modality,and light-weighting has enabled natural language processing to enter a new era of pre-training.This paper mainly summarizes the research methods and research conclusions of lightweight pre-training models,knowledge-incorporated pre-training models,cross-modal pre-training language models and cross-language pre-training language models,as well as the main challenges faced by the pre-training model expansion model.In summary,four research trends for the possible development of extended models are proposed to provide theoretical support for beginners who learn and understand pre-training models.
Optimal Order Acceptance Decision Based on After-state Reinforcement Learning
QIAN Jing, WU Ke-yu, CHEN Chao, HU Xing-chen
Computer Science. 2022, 49 (11A): 210800261-9.  doi:10.11896/jsjkx.210800261
Abstract PDF(3720KB) ( 998 )   
References | Related Articles | Metrics
As the diversification of customer demand increases,the make-to-order(MTO) model,i.e.,adapting production scheme according to customers’ orders,has attracted increasingly more attention from industry.How to determine whether to accept incoming orders according to the limited production capacity and order status of the enterprise,which is crucial for the enterprise to improve profits.On the basis of the traditional order acceptance problems,this paper proposes a more complete model.Besides the traditional model elements(including delayed delivery cost,rejection cost,and production cost),we further consider the order inventory cost,customer priority and others.Moreover,we model the optimal order acceptance problem as a Markov decision process(MDP).In addition,because the classic MDP method relies on solving and estimating high-dimensional state value function,its computation complexity is high.Therefore,in order to reduce the complexity,this paper proves that the optimal strategy based on the state value function in the classical MDP problem can be defined and constructed by the value function based on the after-state equivalent,thus transforming the multi-dimensional control problem into a one-dimensional control problem.At the same time,in order to solve the continuous state space,this paper combines neural network to parameterize the after-state value function,and solves the problem of large state space.Finally,simulation experiments verify the applicability and superiority of the proposed order acceptance strategy model and algorithm.
Sub-BN-Merge Based Bayesian Network Structure Learning Algorithm
ZHONG Kun-hua, CHEN Yu-wen, QIN Xiao-lin
Computer Science. 2022, 49 (11A): 210800172-7.  doi:10.11896/jsjkx.210800172
Abstract PDF(2221KB) ( 1142 )   
References | Related Articles | Metrics
Aiming at the problem that Bayesian network structure learning method K2 requires to provide accurate prior node order information which is difficult to obtain in practice,and hill-climb algorithm is highly dependent on initial network structure and easy to fall into local optimal,.a sub-BN merge based Bayesian network structure learning algorithm,(Sub-BN-Merge) is proposed in this paper.Firstly,for each node,the Sub-BN-Merge algorithm constructs a sub-net and merges them via voting to get a candidate parent set.Next,based on a scoring criterion,the algorithm searches the temporary optimal parent set of every nodes.Then,we breaks the cycle in the directed graph of the search result to get an directed acyclic graph.At last,further optimization is executed via a heuristic search method with directed acyclic graph as the initial value to get the final Bayesian network structure.Experiment is carried out on small network Asia,medium network Alarm and large network Win95pts.we also analyze the algorithm performance for the missing value situation.Experimental results show the effectiveness of the proposed algorithm.Our Sub-BN-Merge algorithm outperforms the comparison algorithms in term of structural Hamming distance and algorithm correct rate.
Gated Two-tower Transformer-based Model for Predicting Antigenicity of Influenza H1N1
LI Chuan, LI Wei-hua, WANG Ying-hui, CHEN Wei, WEN Jun-ying
Computer Science. 2022, 49 (11A): 211000209-6.  doi:10.11896/jsjkx.211000209
Abstract PDF(2629KB) ( 1019 )   
References | Related Articles | Metrics
The rapid evolution of influenza virus hemagglutinin protein has led to the continuous production of new virus strains,which may cause seasonal influenza and even global influenza outbreaks.Timely detection of antigen variants is essential for vaccine screening and design.Therefore,a robust predictive model of antigenicity is an effective method to deal with the challenge of vaccines.Various end-to-end feature learning tools provide good feature representation methods for proteomics,but the existing influenza A prediction models cannot effectively extract and utilize features in amino acid sequences.In this paper,a gated two-tower model is designed based on the transformer.By inputting the amino acid sequence of the influenza A virus hemagglutinin protein,two parallel encoders are used to capture the antigenic characteristics from the time and space dimensions of the hemagglutinin protein amino acid sequence,and learn the nonlinear relationship between features and prediction results.In order to reduce the noise in the data,when fusing the features in the time dimension and the space dimension,the weights that measure their relative importance are adaptively obtained through the gate mechanism for selective fusion,and finally the fusion features are used to predict the H1N1 influenza antigen variants.Experimental results on the H1N1 data set show that the use of the model’sexcellent non-linear feature learning ability improves the predictive performance of antigenic variation,and at the same time has good robustness.
Multi-view Distance Metric Learning with Inter-class and Intra-class Density
REN Shuang-yan, GUO Wei, FAN Chang-qi, WANG Zhe, WU Song-yang
Computer Science. 2022, 49 (11A): 211000131-6.  doi:10.11896/jsjkx.211000131
Abstract PDF(2148KB) ( 1107 )   
References | Related Articles | Metrics
Geometric information can provide prior knowledge and intuitive explanation for classification methods.Observing samples from geometric perspective is a novel method of sample learning,and density is a very intuitive form of geometric information.This paper proposes a multi-view distance metric learning method with inter-class and intra-class density to learn a metric space.In this space,the heterogeneous samples are more scattered,and the homogeneous samples are closer.First,the inter-class density is introduced under the large margin framework,and the samples in the metric space are constrained by minimizing the inter-class density,so as to realize the inter-class dispersion and improve the classification performance.Second,maximize the intra-class density to achieve the effect of similar samples close to each other,so as to achieve intra-class compactness.Finally,to better mine the complementary information of the multi-view samples,the correlation between the views in the metric space is maximized,so that the views can learn from each other adaptively and explore the complementary information among the views.A large number of experimental results on real-world datasets demonstrate the superiority of the proposed method.
Multi-UAV Cooperative Exploring for Large Unknown Indoor Environment Based on Behavior Tree
SHI Dian-xi, SU Ya-qian-wen, LI Ning, SUN Yi-xuan, ZHANG Yong-jun
Computer Science. 2022, 49 (11A): 210900083-11.  doi:10.11896/jsjkx.210900083
Abstract PDF(4482KB) ( 1057 )   
References | Related Articles | Metrics
This paper proposes a method of using behavior tree framework to schedule multiple UAVs and path planning algorithms for collaborative exploration in a large unknown indoor space without GPS signals.The core of this method is to use the Tracking-D*Lite algorithm to track moving targets in unknown terrain,combined with the Wall-Around algorithm based on the Bug algorithm to navigate the UAV in the unknown indoor environment.Finally,the behavior tree is used to schedule and switch multiple UAVs and these two algorithms.This method is based on ROS and uses Gazebo for simulation and visualization.It designs and implements comparative experiments with other unknown indoor environment exploration methods.Experimental results show that it can effectively complete the exploration task and finally draw the boundary contour map of the entire unknown indoor environment.Once extended to the real world,this method can be applied to dangerous buildings after earthquakes,hazar-dous gas factories,underground mines,or other search and rescue scenarios.
Improved Algorithm for Tabu Search of Graph Coloring Problems
WANG Jian-chang, WANG Shuo, LI Zhuang, JIANG Hua
Computer Science. 2022, 49 (11A): 211000128-5.  doi:10.11896/jsjkx.211000128
Abstract PDF(1983KB) ( 1335 )   
References | Related Articles | Metrics
The graph coloring problem is an NP-hard problem,which has a wide range of applications in reality,such as register allocation,airport scheduling and so on.Tabu search algorithm is a classic heuristic search algorithm,which is widely used in the algorithm design of graph coloring problems.As a low-level operator,the tabu search algorithm is also commonly used in the design of graph coloring algorithms such as hybrid evolutionary algorithm(HEA),which plays a key role in the performance of the algorithm.Therefore,the improvement of the tabu search algorithm has practical significance for promoting the research of graph coloring algorithm.Aiming at the problem of graph coloring,this paper proposes an improved version of Tabucol+ to enhance the concentration of search.Based on the traditional tabu search strategy,the proposed algorithm introduces a new scoring strategy.Experimental results show that the new algorithm can significantly reduce the number of iterations andsearch time.In some cases,even the improvement of the number of colors has been achieved.
Personalized Dialogue Generation Integrating Sentimental Information
XU Hui, WANG Zhong-qing, LI Shou-shan, ZHANG Min
Computer Science. 2022, 49 (11A): 211100019-6.  doi:10.11896/jsjkx.211100019
Abstract PDF(2473KB) ( 1091 )   
References | Related Articles | Metrics
Nowadays,more and more attention has been paid to the man-machine dialogue system.However,the current mainstream man-machine dialogue system rarely considers the personalized characteristics of the speaker.An important aspect of the dialogue system is to improve the response quality of dialogue according to the personality of interactive personnel.Personalization is the key to create intelligent dialogue system,which can be well adapted to human life.Emotion is a very important factor in the generation of personalized dialogue.Therefore,a personalized dialogue generation model integrating attribute level emotion is proposed in this paper.The BERT-MRC model is used to extract the emotional and attribute information of character personality and historical dialogue.The improved UNILM neural network model is used to encode character personality and historical dialogue.At the same time,the emotional word information and attribute word information are combined in the coding representation to finally generate a dialogue in line with character personality.Experiments show that the proposed method can effectively improve the quality of personalized dialogue generation and increase the diversity of generated responses.
Online Learning Emotion Recognition Based on Videos
WEI Yan-tao, LUO Jie-lin, HU Mei-jia, LI Wen-hao, YAO Huang
Computer Science. 2022, 49 (11A): 211000049-6.  doi:10.11896/jsjkx.211000049
Abstract PDF(3178KB) ( 1159 )   
References | Related Articles | Metrics
With the normalization of epidemic prevention and control,online learning has become one of the main forms of daily teaching activities.However,with the large-scale development of online learning activities,the problem of “emotional loss” is increasingly prominent,which has become the main reason for the low completion rate of online learning.Aiming to deal with the above problems,the non-invasive online learning emotion state recognition method using video data is discussed.Firstly,the facial videos and heart rate data of 22 students learning online are collected to construct a bimodal online learning emotion database.Secondly,the frame attention network is used to extract facial expression features from the learning video and recognize the emotional state of online learning,and its recognition accuracy reaches 87.8%.Finally,the application of the video heart rate recognition method in online learning emotion analysis is discussed.Research results show that the heart rate level in the confused state is significant.Starting from learning video data mining,focusing on learning emotion recognition based on facial expressions and video heart rate,which provides a new idea for improving the perception of emotional state in online learning.
Study on Dual Sequence Decision-making for Trucks and Cargo Matching Based on Dual Pointer Network
CAI Yue, WANG En-liang, SUN Zhe, SUN Zhi-xin
Computer Science. 2022, 49 (11A): 210800257-9.  doi:10.11896/jsjkx.210800257
Abstract PDF(2931KB) ( 1110 )   
References | Related Articles | Metrics
Due to the uneven utilization of road transportation resources in my country,the supply and demand of trucks and cargo become a hot issue today.In order to maximize the utilization of overall transportation resources,the truck and cargo supply-demand matching platform needs to integrate transportation demand and capacity,reduce costs and improve efficiency.The algorithms used by most platforms are usually heuristic algorithms to solve the problem of trucks-cargo matching.Such algorithms have a bottleneck in optimizing when faced with large-scale problems.In response to the above-mentioned problems,this paper transforms the supply-demand matching problem of vehicles and goods into a double sequence decision-making problem for the first time.Based on this,we study an efficient algorithm that is suitable for today’s vehicle and goods supply-demand matching links.First,a mathematical model of trucks-cargo matching is proposed and the model is abstracted as a double sequence decision problem,and then a dual-pointer-network algorithm is innovatively proposed to solve this problem.The experiment uses the Actor-Critic algorithm as the model training framework to train the dual-pointer-network and evaluates the model.Experiments show that the dual-pointer-network’s vehicle-to-cargo matching solution method is equivalent to traditional heuristic algorithms in small problem scales,and surpasses heuristic algorithms in large problem scales.At the same time,the time consumption is greatlyreduced.
Prediction of Antigenic Similarity of Influenza A/H5N1 Virus Based on Attention Mechanism and Ensemble Learning
WANG Ying-hui, LI Wei-hua, LI Chuan, CHEN Wei, WEN Jun-ying
Computer Science. 2022, 49 (11A): 210900032-6.  doi:10.11896/jsjkx.210900032
Abstract PDF(2012KB) ( 944 )   
References | Related Articles | Metrics
Influenza A virus can lead to seasonal influenza virus outbreaks or even global outbreaks.Continued and cumulative changes in the hemagglutinin protein of influenza viruses can lead to the antigenic variants that reduce vaccine effectiveness or even cause vaccine failure.Therefore,antigenic similarity prediction is critical for influenza outbreak surveillance and vaccine selection.Although A/H5N1 virus originates in poultry,they can cause pneumonia and multiple organ failure in humans.In view of influenza virus and the antigenic characteristics,this paper designs a neural network model to predict the antigenic similarity between viruses.Specifically,the model represents amino acid sequences based on the K-mer embedding and position specific scoring matrices(PSSM),then integrates the features.Furthermore,integrated deep learning model fused with attention mechanism for antigen similarity prediction.Experimental results show that the model significantly improves the accuracy,precision,F1 and MCC compares with the baseline models.Experimental results show that the model has good robustness and extensibility,and has good application potential in the field of antigenic similarity prediction.
Cross-lingual Term Alignment with Kernel-XGBoost
YU Juan, ZHANG Chen
Computer Science. 2022, 49 (11A): 211000111-6.  doi:10.11896/jsjkx.211000111
Abstract PDF(2599KB) ( 1149 )   
References | Related Articles | Metrics
Cross-lingual term alignment is a crucial step for cross-lingual text data analysis and knowledge discovery.Current research usually focuses on single-word term alignment and relies heavily on vector space alignment.Therefore,a new Kernel-XGBoost method is proposed for the one-to-many alignment of cross-lingual terms including multi-word terms.Given a cross-lingual parallel corpus,the proposed method obtains synonymous cross-lingual terms in two steps:1) extracting cross-lingual terms and generating candidate term pairs;2) aligning cross-lingual terms based on word embedding.Experiments on Chinese-Spanish and Chinese-French term alignments demonstrate that the proposed method can achieve an accuracy of 80% at Top-5.It can effectively support cross-lingual text mining tasks such as information retrieval,ontology building.
Stance Detection Based on Argument Boundary Recognition
CHEN Zi-zhuo, LIN Xi, WANG Zhong-qing
Computer Science. 2022, 49 (11A): 210800180-5.  doi:10.11896/jsjkx.210800180
Abstract PDF(3282KB) ( 988 )   
References | Related Articles | Metrics
In online debates,people often use emotional language to make their points with casual expressions.As a supporting evidence of these points,argument contains more of the author’s stance polarity,so the argument information is of great help to the stance detection.In this paper,we propose a stance detection model of BERT via argument boundary recognition to improve the effect of classification on texts in online debates.We use bidirectional encoder representations from transformers(BERT) to screen the arguments related to the topics from the debate posts.Then we combine the argument information and the text information to predict the stance polarity via BERT.Experiments are carried out on the English online debate dataset,compared with the stance analysis model of BERT using only text information and other machine learning models,the model based on argument boundary recognition has better performance in stance detection.
Empirical Study on the Forecast of Large Stock Dividends of Listed Companies Based on DE-lightGBM
CEN Jian-ming, FENG Quan-xi, ZHANG Li-li, TONG Rui-chao
Computer Science. 2022, 49 (11A): 211000017-7.  doi:10.11896/jsjkx.211000017
Abstract PDF(2816KB) ( 926 )   
References | Related Articles | Metrics
Large stock dividends refers to the transfer of a large proportion of shares by listed companies.Aiming at the prediction problem of large stock dividends phenomenon implemented by listed companies,this paper proposes alightGBM based on Differential Evolution algorithm hyperparametric optimization(Named as DE-lightGBM).The model mainly includes two aspects:Firstly,Differential Evolution algorithm is used to adjust the weight of a few categories and the coefficient of regular term in the loss function of lightGBM to deal with the problem of data category imbalance.Secondly,taking F1 and AUC as evaluation indexes,Differential Evolution algorithm is used to optimize the important hyperparametric variables of lightGBM model again to find a group of parameter combinations with the best prediction effect.The numerical results show that the DE-lightGBM has achieved good results,and the F1 and AUC are 0.536 8 and 0.873 4 respectively.DE-lightGBM proposed in this paper can effectively identify the listed companies that will implement stock dividends next year.
Text Classification Based on Knowledge Distillation Model ELECTRA-base-BiLSTM
HUANG Yu-jiao, ZHAN Li-chao, FAN Xing-gang, XIAO Jie, LONG Hai-xia
Computer Science. 2022, 49 (11A): 211200181-6.  doi:10.11896/jsjkx.211200181
Abstract PDF(2970KB) ( 1042 )   
References | Related Articles | Metrics
Text emotion analysis is often used in word-of-mouth analysis,topic monitoring and public opinion analysis.It is one of the most active research fields of natural language processing.The pre-training language model in deep learning can solve the problems of polysemy,part of speech and its position in text emotion classification task.However,its model is complex and has many parameters,which leads to huge consumption of resources and difficult deployment of the model.To solve the above problems,using the idea of knowledge distillation,using ELECTRA pre-training model as teacher model and BiLSTM as student model,a distillation model based on ELECTRA-base-BiLSTM is proposed.The word vector representation encoded by text “one-hot” is used as the input of distillation model to classify Chinese text emotion.Through experimental verification,the distillation results of seven teacher models including ALBERT-tiny,ALBERT-base,BERT-base,BERT-wwm-ext,ERNIE-1.0,ERNIE-GRAM and ELECTRA-base are compared respectively.It is found that ELECTRA-base-BiLSTM distillation model has the highest accuracy,precision and comprehensive evaluation indicators,and the best emotion classification effect,which can obtain text emotion classification results close to ELECTRA language model.Its classification accuracy is 5.58% higher than that of lightweight shallow network BiLSTM model.This model not only reduces the complexity of ELECTRA model and reduces resource consumption,but also improves the effect of Chinese text emotion classification of lightweight BiLSTM model,which has a certain reference value for the subsequent research of text emotion classification.
Multi-target Regression Method Based on Hypernetwork
SUN Kai-wei, GUO Hao, ZENG Ya-yuan, FANG Yang, LIU Qi-lie
Computer Science. 2022, 49 (11A): 211000205-9.  doi:10.11896/jsjkx.211000205
Abstract PDF(3924KB) ( 961 )   
References | Related Articles | Metrics
Multi-target regression(MTR) is a kind of machine learning problem which predicts multiple relevant continuous output targets simultaneously.In MTR,multiple output targets share the same input feature representation,the main challenge of MTR lies in how to effectively explore and utilize the correlations among multiple output targets to improve the prediction accuracy of all output targets.In this paper,a multi-target regression method based on hypernetwork(MTR-HN) is proposed.First,the k-means method is applied to each output target to divide it into multiple clusters.Then,according to the clustering results,MTR problem is transformed into a multi-class multi-label classification problem.Finally,hypernetwork model is utilized to model the multi-class multi-label classification problem,and the final prediction model for MTR is built based on hypernetwork.The main merits of MTR-HN lie in:1)discretizing the output space,can reduce the risk of overfitting;2)hypernetwork can model the inter-target correlations more effectively.Comparative experiments on 18 multi-target regression datasets show that the proposed MTR-HN achieves better prediction performance than existing state-of-the-art multi-target regression methods.
Customs Synonym Recognition Fusing Multi-level Information
LIU Da-wei, CHE Chao, WEI Xiao-peng
Computer Science. 2022, 49 (11A): 210800197-5.  doi:10.11896/jsjkx.210800197
Abstract PDF(2756KB) ( 1120 )   
References | Related Articles | Metrics
In the text information of customs import and export commodities,different words are often used to describe the chara-cteristics of the same commodity.Recognizing the characteristic synonyms of these commodities can better summarize opinions,and then prevent and control the tax-related risks for commodities with the same characteristics.According to the characteristics of phrases of customs declaration elements,a convolution neural network model fusing multi-level information is proposed,and a Sentence-BERT based on twin and three-level network structure is constructed and trained,which has a better semantic representation of similar element phrases,and makes up for the shortage of discrete and sparse embedded features of short text words in Word2Vec.Multi-size convolution kernel is used to extract different features of keywords.The BiLSTM neural network is used to learn the word order information of element phrases,and the attention mechanism is used to assign the weight of keywords.The full connection layer integrates semantic features of synonyms and keyword features,and is predicted by SoftMax layer.Experiments show that the convolution model fusing multi-level information has better performance than other models.
Improved Ant Colony Algorithm for Solving Multi-objective Unilateral Assembly Line Balancing Problem
WU Xiao-wen, ZHENG Qiao-xian, XU Xin-qiang
Computer Science. 2022, 49 (11A): 210900165-5.  doi:10.11896/jsjkx.210900165
Abstract PDF(2069KB) ( 986 )   
References | Related Articles | Metrics
In response to the second type of unilateral assembly line balance in industrial production,it is established with minimal chemical sections under two constraints.In these two constraints,Mathematical model for optimizing the balance loss rate.A feasible solution to the unilateral assembly line balance problem using an improved ant colony algorithm combining ant colony algorithm and simulated annealing algorithm.The algorithm uses the operation selection mechanism and the operation allocation mechanism to select and assign it to a station.In the experiment,simulation examples are used to verify the feasibility of the algorithm,and 8 examples are used to verify the effectiveness of the proposed algorithm for solving the problem,and to produce certain technical method support for actual industrial production.
Rescheduling of Production System Under Interference of Emergency Order
HUANG Peng-peng, ZHAO Chun, GUO Yu
Computer Science. 2022, 49 (11A): 211100193-6.  doi:10.11896/jsjkx.211100193
Abstract PDF(2770KB) ( 1070 )   
References | Related Articles | Metrics
The rescheduling of all orders is investigated after the insertion of emergency order into the mixed-flow production system.First,according to the product’s technology and equipments used,the similarity between the emergency order and the original virtual cells is calculated.Similar rush orders are then inserted into existing cells and the production resource scheduling scheme is adjusted to prioritise the production of rush orders.In order to reduce the impact of rescheduling on the production system and take into account the efficiency and stability of production,a mathematical model is constructed with the goal of minimizing the total process time and product sequence disturbance for all task orders,and a genetic-ant colony algorithm is designed to solve the problem with the positive feedback of the ant colony algorithm after finding a better solution by the genetic algorithm.Finally,an example is substituted into the constructed model and solved with the help of MATLAB programming.The results show that the method could optimize the allocation of production resources for rescheduling and ensure the efficiency and stability of enterprise production.
Multi-armed Bandit Model Based on Time-variant Budgets
LIN Bao-ling, JIA Ri-heng, LIN Fei-long, ZHENG Zhong-long, LI Ming-lu
Computer Science. 2022, 49 (11A): 210800212-6.  doi:10.11896/jsjkx.210800212
Abstract PDF(2161KB) ( 839 )   
References | Related Articles | Metrics
Many budget-related multi-armed bandit models have been proposed,but the practical problems they can solve are limi-ted,that is,they must all be subject to a total budget limit.Therefore,this paper proposes a multi-armed bandit model based on time-variant budgets,which can break this limitation and be used to solve other practical problems.The model captures the situation where the learner’s actions for each round are limited by the corresponding round budget.More specifically,at each round,the player is required to choose to pull L(L≥1) arms(L is not a fixed value) within the budget limits of that round.The player’sgoal is to maximize the total average reward within the budget limits of each round.According to this model,a dynamic programming algorithm based on confidence bound is proposed.The algorithm takes advantage of the characteristics of the model,takes the confidence upper bound of the empirical average reward of the arm as the basis for each round,and then uses the dynamic programming algorithm to perform the arm pull operation.In this paper,the concept of regret is introduced,and it is deduced theoretically that there is a relationship between the upper bound of regret and the sum of budget.Finally,the feasibility of the proposed algorithm is verified by comparing the proposed algorithm with other traditional budget-limited multi-armed bandit algorithms(ε-first,KUBE,BTS) under different scenarios.
EGOS-DST:Efficient Schema-guided Approach to One-step Dialogue State Tracking for Diverse Expressions
ZHU Ruo-chen, YANG Chang-chun, ZHANG Deng-hui
Computer Science. 2022, 49 (11A): 210900246-7.  doi:10.11896/jsjkx.210900246
Abstract PDF(2506KB) ( 863 )   
References | Related Articles | Metrics
Recent dialogue state tracking works have focused on the hybrid approach to balance the two extreme methods(i.e.,over-reliance on ontology and complete abandoning ontology).However,some special phenomena are ignored in these works.For instant,value sharing and recommendation acceptance.In addition,the widely used slot gate mechanism makes it difficult for the model to process slots in parallel and the mechanism also propagates errors to the slot value generation steps.This paper proposes a new hybrid approach that deals with four different phenomena,namely diverse value,unseen value,value sharing and recommendation acceptance.By modifying the candidate value set and the model input,our model can parallelly process slots in one step and no longer depend on the slot gate.Experimental results indicate that the model achieves 57.7% and 59.5% joint goal accuracy on the English dataset MultiWOZ 2.2 and 2.3,respectively,and reaches 68.1% on the Chinese dataset RiSAWOZ,with only 10ms infer time.Finally,the robustness of the model is analyzed.The results on MultiWOZ 2.2 show that the joint target accuracy rate is 55.4% when the recommendation error rate reaches 15%.
Hybrid Particle Swarm Optimization Algorithm Based on Hierarchical Learning and Different Evolution for Solving Capacitated Vehicle Routing Problem
CHEN Ying, HUANG Pei-xuan, CHEN Jin-ping, WANG Zu-yi, SHEN Ying-shan, FAN Xiao-mao
Computer Science. 2022, 49 (11A): 210800271-7.  doi:10.11896/jsjkx.210800271
Abstract PDF(2519KB) ( 964 )   
References | Related Articles | Metrics
The purpose of the vehicle routing problem(VRP) is to search the service route of each vehicle,so as to minimize the sum of driving distances in the case of completing all of the distribution tasks.CVRP,a classical combinatorial optimization problem in operations research,belongs to NP-hard problem and has high theoretical significance and practical value.In order to solve this problem,a hybrid particle swarm optimization algorithm based on hierarchical learning and different evolution(DE-HSLPSO) is proposed.First,the hierarchical learning strategy is introduced and the population particles are divided into three layers according to their fitness values and number of iterations.Secondly,the social learning mechanism is introduced in the evolution of the first two layers of particles,while particles in the third layer carry out differential evolution which effectively increases the diversity of particles,thus expanding the space and jumping out of local optimal.Simulation experiment whose examples are taken from the classical CVRP data sets explores the impact of each part of DE-HSLPSO on the overall performance.It is found that both hierarchical strategy and differential evolution can improve the overall performance of the algorithm.In addition,DE-HSLPSO and other algorithms are tested on seven benchmark examples.With comprehensive comparison of time and optimal solution,the result shows that the solution performance of DE-HSLPSO is better than that of other algorithms.
Study on Evolution of Sentiment-Topic of Internet Reviews with Time in Emergencies
SHI Wei, FU Yue
Computer Science. 2022, 49 (11A): 211000193-6.  doi:10.11896/jsjkx.211000193
Abstract PDF(2557KB) ( 882 )   
References | Related Articles | Metrics
The analysis of sentiment topic evolution is of great value to the control of network public opinion in emergencies.According to the dynamic characteristics of sentiment topics,this paper constructs a sentiment topic model based on LDA,analyzes the evolution of sentiment topics with time through the joint modeling of time,topic and sentiment,deduces the reasoning algorithm based on Gibbs sampling process,and finally puts forward the analysis results of product reviews and microblog emergency data sets,which shows that the joint model has good accuracy and good applicability in the process of time evolution.
Study on Risk Control Model of Selective Ensemble Algorithm Based on Hierarchical Clustering and Simulated Annealing
WANG Mao-guang, JI Hao-yue, WANG Tian-ming
Computer Science. 2022, 49 (11A): 210800105-7.  doi:10.11896/jsjkx.210800105
Abstract PDF(2253KB) ( 949 )   
References | Related Articles | Metrics
Eensemble learning model can effectively solve the problems of single model structure,stability and weak predictive ability.However,due to the complexity of its structure,problems such as low operating efficiency and excessive storage cost often occur.Selective ensemble algorithms are often used to optimize ensemble learning models to solve these problems.The currently proposed selective ensemble algorithm still has the phenomenon of insufficient operating effect and efficiency improvement.In order to make up for these shortcomings,a selective ensemble algorithm based on the stacking ensemble framework is proposed.It mainly uses the agglomerated hierarchical clustering(AHC) algorithm and the metropolis criterion of simulated annealing to select the type and number of base learners.In terms of empirical analysis,domestic and foreign online loan data are used separately to build the model.Experimental results prove that the selective ensemble model of AHC-Metropolis can effectively improve the computational efficiency,predictive ability,stability and generalization ability.It is helpful for regulating the order of the Internet financial industry,assist in financial supervision tasks,and provide an effective basis for establishing our country’s financial risk control management system and guaranteeing national financial security.
Big Data & Data Science
Survey of Community Detection in Complex Network
PAN Yu, WANG Shuai-hui, ZHANG Lei, HU Gu-yu, ZOU Jun-hua, WANG Tian-feng, PAN Zhi-song
Computer Science. 2022, 49 (11A): 210800144-11.  doi:10.11896/jsjkx.210800144
Abstract PDF(3022KB) ( 1741 )   
References | Related Articles | Metrics
Community structure is an important potential feature that exists widely in complex networks.As a key task of network analysis,mining the community structure has important theoretical and practical significance for exploring the potential characteristics,understanding the network organization structure,and discovering the hidden rules and interaction pattern.This paper introduces the background and significance of community detection,and summarizes and combs the methods of community detection from two aspects:static network community detection and dynamic network community detection.Among them,the community detection methods of static network include community detection based on division,community detection based on hierarchical clustering,community detection based on modularity,community detection based on non-negative matrix factorization and community detection based on deep learning.Dynamic network community detection methods include incremental clustering community detection and evolutionary clustering community detection.This paper also introduces the commonly used evaluation metrics of community detection.Finally,some challenges faced by community detection and the future development direction are discussed.
Data Fusion Method of Network Collaborative Manufacturing Based on Rule Chain
HU Chu-yang, LIU Xian-hui, ZHAO Wei-dong
Computer Science. 2022, 49 (11A): 220300175-7.  doi:10.11896/jsjkx.220300175
Abstract PDF(5118KB) ( 999 )   
References | Related Articles | Metrics
The efficient organization and utilization of manufacturing resource data with effective data fusion to extract more beneficial information is a hot research topic in the field of intelligent manufacturing today.Moreover,the rule chain technology based on event flow is emerging,and its high degree of freedom provides the possibility of combining with data fusion.In view of the problems that there are few manufacturing resource organization models,the application scope of data fusion models is governed within the same architecture system,and the combination with rule chains is less studied,we model the organization method of complex manufacturing resources,improve the scheduling process based on MROM-VMC,summarize the manufacturing resource data storage chain structure,and propose a rule chain-based data fusion for the data processing link in the data chain.A rule-based data fusion method is proposed for the data processing link in the data chain to handle a large amount of homogeneous and heterogeneous sensor data,which is finally validated in the data processing link in the network collaborative manufacturing resource platform.It improves the efficiency of manufacturing resource data utilization and the operability of the data fusion method,and provides users with auxiliary decision support.
Ranking and Recognition of Influential Nodes Based on k-shell Entropy
YUAN Hui-lin, FENG Chong
Computer Science. 2022, 49 (11A): 210800177-5.  doi:10.11896/jsjkx.210800177
Abstract PDF(3741KB) ( 1067 )   
References | Related Articles | Metrics
The spreading capacity of nodes has been one of the most attractive problems in the field of complex networks.Due to the large size of nodes in network,researchers want to find accurate measures to estimate the spreading capacity of nodes.In this paper,a new method is proposed based on the basic concepts of information theory and k-shell,which measures the spreading capacity of nodes according to the topological information of their locations in the network.Experimental results show that the proposed method is more effective than other similar methods,and can effectively avoid the “rich club phenomenon” of k-shell method.
Dynamic and Static Relationship Fusion of Multi-source Health Perception Data for Disease Diagnosis
HUO Tian-yuan, GU Jing-jing
Computer Science. 2022, 49 (11A): 211100241-9.  doi:10.11896/jsjkx.211100241
Abstract PDF(3300KB) ( 975 )   
References | Related Articles | Metrics
Disease diagnosis is a field of electronic health record data mining where lots of researchers are interested in,and it is also an important link to realize the intellectualization of medical diagnosis.However,due to the diversity of data sources,complex data structure and potential correlation among different types of health sensing data,there is a problem of how to fuse heterogeneous data in the process of feature extraction and data mining.Therefore,comprehensively considering clinical sensing data,personal physical record data and relationship data between diseases,and mining the latent relevant features can make the diagnosis of multi-category diseases more accurate.Dynamic and static relationship fusion of multi-source health perception data for disease diagnosis(DSRF) is proposed.Firstly,the dynamic and static relationship fusion algorithm is used to extract data correlation features and solve the heterogeneity of dynamic clinical sensing time series data and static personal physical condition data.Then the dependency matrix of multi-category diseases is calculated to extract the correlations among diseases.Finally,various health sen-sing data is fused based on the gated recurrent unit network.The comprehensive analysis of multi-source heterogeneous data is completed after the above three steps.Experimental results on the real-world American MIMIC-III clinical dataset show that the proposed model outperforms state-of-the-art models and is able to diagnose multiple categories of diseases accurately.
Community Discovery Method Based on Influence of Core Nodes
YUAN Hui-lin, HAN Zhen, FENG Chong, HUANG Bi, LIU Jun-tao
Computer Science. 2022, 49 (11A): 211100002-7.  doi:10.11896/jsjkx.211100002
Abstract PDF(3274KB) ( 935 )   
References | Related Articles | Metrics
Community discovery is a hot topic in the field of complex networks.Many local community detection algorithms have been proposed to quickly discover high-quality communities,but most of them have seed-dependent or stability problems.Some algorithms try to accurately find the seed nodes according to the topology characteristics of the core nodes that they are highly surrounded by neighbors and far away from each other to avoid the above problems.But the calculation of distance makes its time complexity is high.In this paper,a community detection method based on influence of core nodes(CDIC) is proposed.This me-thod first searches for all possible core nodes according to the topological characteristics of core nodes and network adjacency information.Then it uses the higher influential of true core nodes and the idea of label propagation to expand the communities and eliminate nodes wrongly selected as the core to avoid the seed-dependent problems.Besides,the calculation without distance also ensures low time complexity.Finally,a community attraction to nodes based on the similarity theory is proposed to merge specific nodes to ensure the stability of the results.The normalized mutual information and purity of the proposed method,6 classic algorithms and 2 algorithms proposed in recent years are compared on 64 artificial networks and 4 real networks.The results show the effectiveness of CDIC.
Mining Spatial co-location Pattern with Dominant Feature
XIONG Kai-fang, CHEN Hong-mei, WANG Li-zhen, XIAO Qing
Computer Science. 2022, 49 (11A): 211000126-7.  doi:10.11896/jsjkx.211000126
Abstract PDF(3167KB) ( 920 )   
References | Related Articles | Metrics
A spatial co-location pattern is a subset of spatial features whose instances frequently locate together in the neighborhood.Traditional co-location pattern does not distinguish the importance of features in the pattern,and ignores the dominant relationship among features.The co-location pattern with dominant feature considers the inequality of features in the pattern,and analyzes the dominant relationship among features,which can be used in many applications.However,the existing methods for mining co-location pattern with dominant feature do not comprehensively consider the possible tendency and influence intensity of one feature dominating other features from the perspective of features’ instances distribution,so that the dominant relationship among features is not properly revealed.This paper first analyzes the spatial distribution of features’ instances in a co-location pattern,proposes the pattern dominance index to measure the possible tendency of a feature dominating other features in a pattern,and proposes the dominant influence index to measure the influence intensity of the dominance tendency.Based on the two new measures,the dominant feature mining of co-location pattern is proposed.Then an efficient algorithm for mining co-location pattern with dominant feature is proposed by optimizing the calculation of new measures.A large number of experiments on real data sets and synthetic data sets verify that the proposed method can effectively identify the dominant feature in a co-location pattern,and it can efficiently mine co-location patterns with dominant feature.
Novel College Entrance Filling Recommendation Algorithm Based on Score Line Prediction andMulti-feature Fusion
WANG Ze-qing, JI Sheng-peng, LI Xin, ZHAO Zi-xuan, WANG Peng-xu, HAN Xiao-song
Computer Science. 2022, 49 (11A): 211100266-7.  doi:10.11896/jsjkx.211100266
Abstract PDF(2646KB) ( 1859 )   
References | Related Articles | Metrics
In recent years,as the number of high school graduates growing,the demand of college entrance filling is increasing.But faced with massive amounts of college entrance data,students always cannot make reasonable decisions conform to their own will in a short time,resulting in filling accident.To address this issue,on the basis of crawling college entrance history data by web spider,a novel college entrance filling recommendation algorithm based on score line prediction and multi-feature fusion(Reco-PMF) is proposed.Firstly,BP neural network is applied to predict all the colleges admission lines of current year.Then,combining with colleges’ rankings,an admission probability algorithm is constructed on the basis of three score related features.Genetic algorithm is employed to optimize the weights of above features.On this basis,recommendation-score is defined to measure admission risk.Finally,a college filling list with multi-admission risk is generated.Experiment results show that,the college admission line prediction algorithm based on BP neural network performs better than other algorithms under all error bounds.Compared with existing on-line services of Baidu and Kuake,Reco-PMF increases the acceptance rates by 14.8% and 24.1%,and improves the average ranking of recommended colleges by 99 and 87 in accepted colleges.
Memory-augmented GAN-based Anomaly Detection
ZHOU Shi-jin, XING Hong-jieHebei
Computer Science. 2022, 49 (11A): 211000202-9.  doi:10.11896/jsjkx.211000202
Abstract PDF(3189KB) ( 1115 )   
References | Related Articles | Metrics
In the training stage of the generative adversarial networks(GAN) based anomaly detection method,its training set consists of only normal data.When training data are sufficient,the GAN based anomaly detection method may obtain smaller reconstruction error.However,in the testing stage,the difference between the reconstruction errors of normal data and those of part novel data is too small,which makes the discriminant performance of the GAN based anomaly detection method become poor.To solve this problem,a memory-augmented GAN based anomaly detection method is proposed.A memory-augmented mo-dule is introduced into the proposed method to make it remember the characteristic of normal data.Hence,the reconstruction error of novel data becomes larger and thus the discriminant ability of the proposed method is enhanced.In comparison with the related approaches,experimental results on MNIST,Fashion-MNIST and CIFAR-10 verify that the proposed method has better detection performance.
DCPFS:Distributed Companion Patterns Mining Framework for Streaming Trajectories
ZHANG Kang-wei, ZHANG Jing-wei, YANG Qing, HU Xiao-li, SHAN Mei-jing
Computer Science. 2022, 49 (11A): 211100268-10.  doi:10.11896/jsjkx.211100268
Abstract PDF(3802KB) ( 1031 )   
References | Related Articles | Metrics
The widespread use of location technology leads to huge volumes of spatio-temporal data collected in the form of tra-jectory data streams.How to discover useful information from it has attracted more and more scholars’ attention.Mining companion pattern from trajectory stream refers to discovering groups with highly similar behaviors at the same time,which is essential for real-time applications of traffic management and recommendation systems.However,the existing research only achieves a second-level response,and it is difficult to respond quickly in milliseconds to large-scale trajectory data.Therefore,this paper proposes a distributed companion patterns mining framework DCPFS.The main contents of our work include:1) In order to reduce the time consumption of the density-based clustering algorithm DBSCAN for large-scale data,this paper proposes a data partition strategy and clustering merging algorithm based on a distributed deployment plan to ensure clustering parallelism and accuracy.2) Because the trajectory movement is directional in reality,we increase the direction dimension to reduce the redundancy in the clustering.3) We designed a parallel intersection algorithm to improve the efficiency of the intersection of clustering results in the pattern mining stage.4) We implement DCPFS on the Flink distributed big data processing platform and use Chengdu taxi GPS dataset and Google life dataset for experiments.Comprehensive empirical study demonstrates that the proposed framework has faster response speed than the baseline method.
Intelligent Operation Framework for Relational Database Application
JIANG Zong-lin, LI Zhi-jun, GU Hai-jun
Computer Science. 2022, 49 (11A): 211200030-9.  doi:10.11896/jsjkx.211200030
Abstract PDF(4556KB) ( 972 )   
References | Related Articles | Metrics
Relational database refers to the database that uses relational model to organize data.It needs to use structured query language(SQL) to operate the data.When facing the application,it can only operate the database according to the program rules set by the developer.The process of modifying and adding program rules is cumbersome and requires a certain degree of professionalism,which is not friendly to ordinary users.In order to improve the expansion and universality of the application of rela-tional database,this paper uses the knowledge representation theory to model the knowledge related to database operation,uses the knowledge representation method of framework combined with rules to establish the general paradigm of database operation,and studies and designs the reasoning algorithm based on the operation characteristics of relational database and the syntax and semantics of relational model and structured query sentences.The user oriented database operation related things are abstracted into logical symbols,and the internal relationship between them is abstracted into the rule constraints between logical symbols.The problem is solved according to the rule constraints represented by logical symbols using the solving system.Based on the above theories and algorithms,a relational database operation framework integrating knowledge representation is designed and implemented,user input is converted into database operation statements to realize database system operation.It can be seen from the application example that the proposed operation framework can be embedded into the application system on the basis of friendly compatibility with relational database.The program rules are easy to expand,the application system has low difficulty in use,update and maintenance,and has strong self adaptability.It can provide users with more flexible and intelligent database ope-ration management and control services.
Novel Method Based on Graph Attentive Aggregation for POI Recommendations
CAI Guo-yong, CHEN Xin-yi, WANG Shun-jie
Computer Science. 2022, 49 (11A): 210800149-5.  doi:10.11896/jsjkx.210800149
Abstract PDF(2162KB) ( 1222 )   
References | Related Articles | Metrics
For services on location-based social network(LBSNs),effective point of interest(POI) recommendation has great economic and social utility.However,how to comprehend the position,structure and behavior related information of LBSNs and proceed reasoning for POI recommendation is still a challenge task.To exploit the heterogeneous information on LBSN,a novel graph attentive aggregation model for POI recommendation(POIR-GAT) is proposed,which exploits both users’ social information and POIs’ attributed information.Firstly,POIR-GAT uses social relationship to construct user-user graph,and extracts user feature vector together with user-POI interaction graph.Secondly,it constructs feature matrix based on different attributes of POIs,obtains hidden factors through matrix decomposition,integrates multiple features into POI feature vector,and learns their common influence on user behavior.Finally,it realizes the integration of social factors and POI features recommended model.Extensive experiments on two public datasets show that the proposed POIR-GAT model can effectively integrate users’ social information and POI feature information,and improve the quality of POI recommendation.
Movie Recommendation Model Based on Attribute Graph Attention Network
SUN Kai-wei, LIU Song, DU Yu-lu
Computer Science. 2022, 49 (11A): 211100106-8.  doi:10.11896/jsjkx.211100106
Abstract PDF(3128KB) ( 947 )   
References | Related Articles | Metrics
In recent years,graph network has been widely used in the field of recommendation and made a great progress.How-ever,the existing methods tend to focus on the interaction modeling of user projects,so the performance is limited by the problem of data sparsity.Therefore,this paper proposes a movie recommendation model based on graph attention network of attribute graph by using additional attribute information.Firstly,an attention-based GNN is proposed,which uses explicit feed-back to calculate the attention score between entities and attributes.Compared with the aggregation method using Laplace matrix,it can distinguish the influence of different attributes on entities more effectively,and the information aggregation between attributes and entities can be more effective.In addition,different entities are affected differently by attributes and behaviors,a fine-grained pre-ference fusion strategy is designed in this paper to calculate a set of preference fusion weights for each entity to make the embedding representation of entities more accurate and personalized.Experimental results on real data set show that the recommendation method that makes full use of attribute information contained in attribute graph can effectively alleviate the problem of data sparsity and is significantly better than other basic algorithm in terms of recall rate and nDCG,two evaluation indexes of movie recommendation.
Fuzzy Multiple Kernel Support Vector Machine Based on Weighted Mahalanobis Distance
DAI Xiao-lu, WANG Ting-hua, ZHOU Hui-ying
Computer Science. 2022, 49 (11A): 210800216-5.  doi:10.11896/jsjkx.210800216
Abstract PDF(1789KB) ( 960 )   
References | Related Articles | Metrics
Fuzzy support vector machine(FSVM) effectively distinguishes the importance of different samples by introducing fuzzy memberships,which reduces the sensitivity of traditional support vector machines to noise data.The membership function designed based on Euclidean distance ignores the overall distribution of samples and does not consider the different importance of sample features.A fuzzy support vector machine method based on weighted Mahalanobis distance is proposed.This method first applies the Relief-F algorithm to estimate the weight of each feature.Then it utilizes the weight for calculating the weighted Mahalanobis distance between the sample and the center of its class.Finally,the fuzzy membership of the sample is calculated based on weighted Mahalanobis distance.Furthermore,considering the difficulty of determining the kernel function and its parameters,a fuzzy multi-kernel support vector machine(FMKSVM) based on weighted Mahalanobis distance is put forward,which combines FSVM with multiple kernel learning methods.The multi-kernel is constructed in the form of weighted sum,and the weight of each kernel is calculated according to the central kernel alignment method(CKA).The proposed method not only reduces the influence of weakly relevant features on classification results,but also enables a more adequate and accurate representation of the data.Experimental results show that,FSVM based on weighted Mahalanobis distance has higher classification accuracy than FSVM based on Euclidean distance and Mahalanobis distance,and the classification performance of FMKSVM based on weighted Mahalanobis distance is superior to that of the single-kernel model.
Recommendation Algorithm Based on Apriori Algorithm and Improved Similarity
DONG Yun-xin, LIN Geng, ZHANG Qing-wei, CHEN Ying-ting
Computer Science. 2022, 49 (11A): 211000005-5.  doi:10.11896/jsjkx.211000005
Abstract PDF(2228KB) ( 984 )   
References | Related Articles | Metrics
In order to alleviate the data sparse problem and improve the accuracy of collaborative filtering algorithm,a recommendation algorithm based on Apriori algorithm and improved similarity is presented.Firstly,it uses Apriori algorithm to mine the potential connections between users,and uses the confidence of the association rules between users to represent the potential connections between users,then constructs a user confidence matrix to fill the user rating matrix.Secondly,the algorithm uses the confidence matrix to improve the traditional similarity calculation formula and build a comprehensive similarity calculation formula between users.Finally,the algorithm uses the filled user rating matrix and the comprehensive similarity between users to make recommendations for users.The proposed algorithm has higher algorithm accuracy than traditional algorithms.Compared with other algorithms,the proposed algorithm can effectively alleviate the long tail problem of the recommendation system,so as to further improve the recommendation quality of the recommendation system.
Distribution Reduction in Fuzzy Order Decision Data Sets with Attention Degree
XU Wei-hua, ZHANG Jun-jie, CHEN Xiu-wei
Computer Science. 2022, 49 (11A): 210700191-5.  doi:10.11896/jsjkx.210700191
Abstract PDF(1759KB) ( 817 )   
References | Related Articles | Metrics
With the advent of the era of big data,the structure of data becomes more and more complex,and the dimensions of data set become higher and higher,which will affect the efficiency of data mining greatly.Therefore,it is necessary to perform data compression or attribute reduction to information systems,that is,to remove unnecessary redundant attributes,reduce data dimensions,and improve the efficiency of data mining.The reduction methods proposed by many scholars in the past regard each attribute as equally important.But in real life,people’s attention to each conditional attribute in the data set is often different.Aiming at this phenomenon,based on the classical fuzzy decision data set,this paper weights different conditional attributes,defines the weighted score function,and further establishes the fuzzy order decision information system with attention degree.Then the distribution function is introduced into the system and the distribution reduction method is established by the distribution discer-nible matrix.Finally,the feasibility of the method is verified by a case study.
Fuzzy Random Events and Its Probabilities Based on Axiomatic Fuzzy Sets
XIE Jian-xiang, PAN Xiao-dong, ZHANG Bo
Computer Science. 2022, 49 (11A): 211100242-7.  doi:10.11896/jsjkx.211100242
Abstract PDF(1828KB) ( 988 )   
References | Related Articles | Metrics
This paper studies the probability of fuzzy random events based on axiomatic fuzzy set,defines fuzzy random events and their corresponding probabilities,discusses some basic properties of probabilities of fuzzy random events,gives the product rule of fuzzy random events probability,and proves the law of total probability of fuzzy random events and Bayes’rule.
Attribute Reduction Algorithm Based on a New q-rung orthopair Fuzzy Cross Entropy
WANG Zhi-qiang, ZHENG Ting-ting, SUN Xin, LI Qing
Computer Science. 2022, 49 (11A): 211200142-6.  doi:10.11896/jsjkx.211200142
Abstract PDF(2078KB) ( 905 )   
References | Related Articles | Metrics
Entropy is an important means to describe the degree of uncertainty of fuzzy sets.In order to reflect the ambiguity produced by the comparison between the membership degree and the non-membership degree in the q-rung orthopair fuzzy set,a related score function is first proposed.Taking into account the shortcomings of the similarity measures of most of the current q-rung orthopair fuzzy sets,a q-rung orthopair fuzzy set cross entropy is proposed,which is more in line with people’s intuition.As there is relatively little research on attribute reduction of q-rung orthopair fuzzy information system at present,through property discussion and theoretical proof,it is found that this kind of cross entropy can be better applied to attribute reduction of q-rung orthopair fuzzy information system.The related attribute reduction algorithm is presented,and an example is given to illustrate the rationality of this algorithm.Secondly,a method to convert ordinary information system into q-rung orthopair fuzzy information system is given.Finally,the rationality and effectiveness of this method are verified by calculating multiple databases in UCI,which provides new ideas for q-rung orthopair fuzzy information system data preprocessing.
Fuzzy Rough Sets Model Based on Fuzzy Neighborhood Systems
RAN Hong, HOU Ting, HE Long-yu, QIN Ke-yun
Computer Science. 2022, 49 (11A): 211100224-5.  doi:10.11896/jsjkx.211100224
Abstract PDF(1773KB) ( 945 )   
References | Related Articles | Metrics
For fuzzy neighborhood systems,upper and lower fuzzy rough approximation operators based on general fuzzy logic operators are proposed,and the basic properties of the operators are investigated.Then,the concepts of neighborhood system of serial,reflexive,symmetric,unary and Euclidean are extended to fuzzy neighborhood systems.Finally,the related algebraic structures of fuzzy rough approximation operators are discussed when the fuzzy neighborhood system is serial,reflexive,symmetric,unary and Euclidean.
Effective Low-frequency Path Mining Method for Information Flow of Networking Information-centric System of Systems
LIN Wen-xiang, LIU De-sheng
Computer Science. 2022, 49 (11A): 211000001-6.  doi:10.11896/jsjkx.211000001
Abstract PDF(3411KB) ( 872 )   
References | Related Articles | Metrics
With the rapid development of information technology and network technology and their widespread use in military field,networking information-centric system of systems comes into being.The networking information-centric system of systems is dominated by information,its main manifestation is the information activity process.The rationality and efficiency of the information activity process directly affect the operational effectiveness of information in the combat system.The use of process mi-ning technology to discover information activity process models from information activity event logs can provide effective support for modelling,testing and enhancement of information activity processes.Simply filtering noise in logs through event frequency analysis can easily lead to the loss of valid low-frequency paths and reduce the accuracy of the mined information activity processes.Combining the special characteristics of military information activities and the effectiveness characteristics of information transfer,a structure aggregation degree based effective low frequency path mining algorithm is proposed.Simulation analysis shows that the method can successfully separate log noise and effective low frequency paths,which is important for mining real and effective information processes.
Image Processing & Multimedia Technology
Survey of Deep Learning Networks for Video Recognition
QIAN Wen-xiang, YI Yang
Computer Science. 2022, 49 (11A): 211200025-10.  doi:10.11896/jsjkx.211200025
Abstract PDF(2351KB) ( 1271 )   
References | Related Articles | Metrics
Video recognition is one of the most important tasks in computer vision research,which is concerned by many resear-chers.Video recognition refers to extracting the key features from different video clips,analyzing these features,and classification of the video.Compared to a single,static picture,there are many significant differences between frames of a video clip.How to tell the differences through the dimension of spatial-temporal information from video clips are well concerned by researchers.Taking video recognition technology as the target of the research,first,this paper introduces the basic concepts of video recognition and challenges in this area,together with some of the most frequently used datasets in video recognition tasks.Then,the classic video recognition methods based on spatio-temporal interest points,dense trajectories,and improved dense trajectories are reviewed.Also,the deep learning network frameworks for video recognition proposed in recent years are then summarized.They are summarized according to the time order of their proposal and grouped by the different architecture of their network.Among them,the video recognition framework based on 2D convolution neural network is introduced,including two-stream convolutional network architecture,long short-term memory network,and long-term recurrent convolutional network.Then,a framework based on a 3D convolutional neural network is introduced,including Slowfast Network,X3D(eXpand 3D) Network.Following that,the pseudo-3D convolutional neural network is introduced,including R(2+1)d network,Pseudo-3D residual network,and a set of light-weight networks based on building models on temporal information.At last,a Transformer-based network is introduced,including Timesformer,video vision Transformer,shifted window Transformer(Swin Transformer).The evolution of these deep learning frameworks,their implementation details and characteristics are analyzed.The performance of each network on different datasets is evaluated,and the applicable scenarios of each network are analyzed.In the end,the future research trend of video recognition network framework is prospected.Video recognition task can automatically and efficiently recognize the category to which the video belongs,and video recognition based on deep learning has a wide range of practical value.
Overview of 3D Reconstruction of Indoor Structures Based on Point Clouds
REN Fei, CHANG Qing-ling, LIU Xing-lin, YANG Xin, LI Ming-hua, CUI Yan
Computer Science. 2022, 49 (11A): 211000176-11.  doi:10.11896/jsjkx.211000176
Abstract PDF(3577KB) ( 1812 )   
References | Related Articles | Metrics
The 3D reconstruction of indoor structure is essentially a multi-task problem of restoring the indoor layout,which can further reconstruct and semantically segment wall details and furniture.This paper mainly introduces the 3D reconstruction of indoor structure based on point cloud data.Firstly,the data set commonly used for 3D reconstruction of indoor structure is summarized,and then the main methods of 3D reconstruction of indoor structure based on point cloud are described and discussed,and the advantages and disadvantages of the three types of reconstruction methods are analyzed and summarized.Finally,the difficulties and challenges faced by the current 3D reconstruction research of indoor structures are explained,and the future research trends are prospected.It can be concluded that the diversity of scenes and tasks completed by most reconstruction models at pre-sent is relatively poor,and the multi-task coordination scheme that uses redundant information from different angles to jointly optimize has great potential in the reconstruction of indoor structures.In addition,the model still needs to be improved for the seamless integration of the indoor and outdoor environments and the full performance of the interior and exterior buildings.
Review on Classification of Breast Cancer Histopathological Images Based on Convolutional Neural Network
ZHANG Xi-ke, MA Zhi-qing, ZHAO Wen-hua, CUI Dong-mei
Computer Science. 2022, 49 (11A): 210800232-9.  doi:10.11896/jsjkx.210800232
Abstract PDF(2708KB) ( 1332 )   
References | Related Articles | Metrics
Histopathological examination of breast cancer is the “gold standard” for the diagnosis of breast cancer.The classification of breast cancer histopathological images based on convolutional neural network has become one of the research hotspots in the field of medical image processing and analysis.Automatic and accurate classification of breast cancer histopathological images has important clinical application value.Firstly,two public datasets widely used in the classification of breast cancer histopathological images and their evaluation criteria are introduced.Then,the research progress of convolutional neural network on two datasets is mainly elaborated.In the process of describing the research progress,the reasons for the low accuracy of some models are analyzed,and some suggestions are given to improve the performance of the models.Finally,existing problems and future prospects of breast cancer histopathological image classification are discussed.
Pedestrian Detection Optimization Method Based on Data Enhancement and SupervisedEqualization in Fisheye Image
SI Shao-feng, ZHANG Sai-qiang, LI Qing, CHEN Ben-yao
Computer Science. 2022, 49 (11A): 210900070-6.  doi:10.11896/jsjkx.210900070
Abstract PDF(3277KB) ( 962 )   
References | Related Articles | Metrics
In recent years,due to the fisheye camera is widely used in the field of intelligent monitoring,many scholars propose pedestrian detection algorithm for fisheye image.However,the fisheye scene is complex and distorted,which leads to the imba-lance of data set sample distribution and algorithm supervision allocation,which will reduce the performance of the detector.To solve these problems,a data enhancement method for pedestrian detection in fisheye image is proposed,which consists of pattern sampling enhancement and angle histogram enhancement.The pattern sampling enhancement focuses on the mining of difficult samples in fisheye image,and the generated new samples enrich the pedestrian patterns near the center of fisheye image.Angle histogram enhancement is based on the idea of histogram equalization,which smooths the angle distribution of fisheye image samples to alleviate the over fitting problem of single scene.In addition,based on Anchor-free fisheye image pedestrian detector,the fusion of location quality prediction and supervision information is proposed to extend Focal Loss to continuous domain to optimize the supervision allocation of detector location branches.Experimental results show that the proposed data enhancement algorithm can effectively alleviate the uneven distribution of fisheye image data set,and show good results in both Anchor-Based and Anchor-Free detectors.The continuous Focal Loss optimization localization supervision method improves the overall performance by 3.8% without increasing the reasoning complexity of the Anchor-Free detector.
Classification Algorithm of Diabetic Retinopathy Based on Attention Mechanism
SUN Fu-quan, ZOU Peng, CUI Zhi-qing, ZHANG Kun
Computer Science. 2022, 49 (11A): 211000213-5.  doi:10.11896/jsjkx.211000213
Abstract PDF(2986KB) ( 1062 )   
References | Related Articles | Metrics
Diabetic retinopathy is one of the important complications of diabetes and the main cause of blindness in the working population.The gap between retinal images is small and easy to be confused.Due to insufficient medical resources and lack of experienced ophthalmologists,it is difficult to carry out large-scale retinal image screening.Therefore,a classification algorithm for diabetic retinopathy based on attention mechanism is proposed to achieve accurate classification of the degree of retinal image lesions.The preprocessing operations such as data enhancement and image enhancement are carried out on data set.Using EfficientNetV2 as the backbone classification network,the attention mechanism is added to the network for fine-grained classification of retinal images,and the transfer learning strategy is used to train the network.The classification accuracy and the second weighted Kappa value of the proposed model are 97.8% and 0.843 respectively,which can effectively classify the disease degree of retinal images.Compared with other models,it has advantages and is of great significance for the diagnosis and treatment of diabetic retinopathy.
MRI and PET/SPECT Image Fusion Based on Subspace Feature Mutual Learning
ZHANG Ying, NIE Ren-can, MA Chao-zhen, YU Shi-shuang
Computer Science. 2022, 49 (11A): 211000171-6.  doi:10.11896/jsjkx.211000171
Abstract PDF(3470KB) ( 1045 )   
References | Related Articles | Metrics
In medical imaging,MRI images provide detailed texture information and better resolution,while PET/SPECT images retain molecular activity information and color function information,so fusing them is an important task.Most of the existing methods have some problems in the fusion process,such as color distortion,blur and noise.Therefore,a new subspace attention-siamese auto-encoding network(SSA-SAEN) is proposed to fully fuse meaningful information from MRI and PET/SPECT images.SSA-SAEN is proposed in image fusion network,and the subspace feature mutual learning concept is introduced.By using subspace attention module,MRI and PET/SPECT images can learn each other’s features,while reducing information redundancy and ensuring efficient and complete feature extraction.In addition,the conditional probability model is used to complement and fuse the extracted features,and the weighted fidelity gradient loss term is added into the training network to achieve the goal of network optimization.A large number of qualitative and quantitative experiments on public datasets show that the proposed me-thod can obtain a clear fused image,which demonstrates the superiority and effectiveness of the proposed method compared with other advanced methods.
Improved Water Quality Remote Sensing Monitoring Algorithms Based on Multilayer Convolutional Neural Network
FENG Lei, FENG Li, FANG Fang, GUO Jin-song, PAN Jiang, YU You, CHEN Yu
Computer Science. 2022, 49 (11A): 210200160-5.  doi:10.11896/jsjkx.210200160
Abstract PDF(2979KB) ( 1057 )   
References | Related Articles | Metrics
With the rapid development of water environment online monitoring technology in recent years,the categories and quantities of monitoring data have been greatly improved.Online monitoring and remote sensing monitoring are important data sources for water environment monitoring.How to quickly and efficiently understand massive monitoring data is a research hot-spot of artificial intelligence technology in the field of ecological environment data research.Changshou lake is a national good water body in the Three Gorges reservoir area.This paper aims at proposing an improved CNN convolution neural network algorithm WRCNN model,and this model is studied to extract features directly from water environment data in remote sensing images and increasing data dimension of water monitoring data.The ability of extraction can eliminate the uncertainty of function selection,reduce the calculation steps,suppress the influence of over-fitting and realize the application of remote sensing technology of large-scale monitoring in water environment.The results show that the improved WRCNN convolution neural network algorithm model can effectively identify the concentration of chlorophyll,the indicator of eutrophication in Changshou lake,and provide an efficient measures for monitoring eutrophication in reservoir area.
Study on Human Pose Estimation Based on Multiscale Dual Attention
MA Wan-yi, ZHANG De-ping
Computer Science. 2022, 49 (11A): 220100057-5.  doi:10.11896/jsjkx.220100057
Abstract PDF(2817KB) ( 940 )   
References | Related Articles | Metrics
In view of the problem of low discrimination between human body and background in human posture estimation,and incomplete utilization of important feature information in human posture estimation based on HRNet,a human posture estimation method MDA-HRNet based on multiscale dual attention is proposed by using channel and spatial attention mechanism.Conside-ring both of the channel domain and spatial domain,the Ca-Neck and Ca-Block modules combined with channel attention and Sa-Block module combined with spatial attention are designed respectively.Then integrating these modules into the high-resolution network structure,so that the network can pay more attention to the human body area in the image.Moreover,in the Sa-Block module,3×3 and 7×7 convolution kernels are adopted to derive two spatial attention maps of different scales,which makes the ability of the network to comprehensively distinguish human features and background features more remarkable,so as to accurately locate the human body and its key points.The proposed method is tested and verified on MPII data set,and the results show that MDA-HRNet can improve the accuracy of joint point location of human posture estimation effectively.
Static Gesture Recognition Based on OpenCV in Simple Background
XU Yue, ZHOU Hui1 School of Computer Science, Technology, Xi’an Jiaotong University, Xi’an 710049, China
Computer Science. 2022, 49 (11A): 210800185-6.  doi:10.11896/jsjkx.210800185
Abstract PDF(3000KB) ( 1063 )   
References | Related Articles | Metrics
Gesture recognition is a very important technology in human-computer interaction,which has high theoretical and practical value.However,due to the complexity of the background and individual differences,gesture recognition has become a challenging topic.Therefore,it is necessary to design an efficient and accurate gesture recognition algorithm to effectively recognize the detected target gesture.An improved method of gesture segmentation and gesture feature extraction is proposed.SVM classifier is used to construct gesture model and recognize gesture.On the basis of YCrCb color space,OTSU threshold processing method is combined to select threshold segmentation gesture to improve the accuracy of segmentation.On the basis of edge detection,the ellipse Fourier descriptor is used to fit the edge and extract gesture features.Experimental results show that the system based on the above algorithm can extract gesture feature information efficiently,and the average recognition accuracy of 13 common gestures in a simple background is 89.96%,which can basically meet the requirements of recognition accuracy and stability.
Using Image Stabilization and VIBE Algorithm to Extract Foreground Target from Jitter Video
LIU Xin-fu, JIANG Mu-rong, HUANG Ya-qun, ZHANG Zhan-wei, ZHOU Ying-ying
Computer Science. 2022, 49 (11A): 210800195-8.  doi:10.11896/jsjkx.210800195
Abstract PDF(6642KB) ( 1004 )   
References | Related Articles | Metrics
In target detection and tracking processing area,the extracted foreground target is often disturbed by background jittering,resulting the problems of target misjudgment,incomplete contour and so on.In this paper,we present a method to improve the integrity of foreground target extraction by combining image stabilization and frame supplement,and then using VIBE algorithm and connected domain correction.Firstly,the image of two adjacent frames is matched by ORB feature,and the wrong ma-tching points are eliminated by RANSAC method.Then,the image margin is adjusted with the moving target as the center,and the non-coincident parts are cropped and repaired.The offset mean value of two adjacent frames is calculated by statistical method to achieve the purpose of video image stabilization.Secondly,the dense optical flow method is used to calculate the displacement of two adjacent frames,and then the intermediate frame is generated by position remapping and supplemented to make the video sequence more smooth and stable.Thirdly,based on VIBE algorithm,morphological processing and connected domain correction are added,and Canny operator is combined for edge detection to increase the integrity of target contour.Finally,a video example is used to test and compare with the commonly used video target extraction algorithms.The precision and recall rates increase by at least 10%.The rates of FNR and PWC reduce by at least 15%.Experimental results show that the proposed method can effectively remove the influence of jitter in most dynamic backgrounds,and the integrity of the extracted target is high.
Classification Method of Harmful Garbage Images Based on Improved EfficientNetV2
YUAN Hui-lin, LIU Jun-tao, HUANG Bi, HAN Zhen, FENG Chong
Computer Science. 2022, 49 (11A): 211100100-5.  doi:10.11896/jsjkx.211100100
Abstract PDF(2894KB) ( 1385 )   
References | Related Articles | Metrics
With the rapid development of industry,the amount of waste generated has also exploded,making waste disposal a worldwide problem.The Chinese government’s concern for the environment has gradually deepened,and various garbage classification policies and laws and regulations have been continuously introduced to supervise citizens’ garbage classification.Garbage disposal,especially hazardous garbage such as electronic waste,if it is improperly handled,will result in bad influence.Hazardous garbage image data has the characteristics of low data quality and unclear images.The image samples collected from different devices have obvious differences.Therefore,the image processing of hazardous garbage faces huge challenges.At the same time,the classification results of hazardous waste are related to environmental pollution issues.The amount of waste produced is huge,requiring high processing speed and accuracy.This paper proposes a garbage image classification method based on convolutional neural network and hybrid attention mechanism.This method does not need to manually extract features from the input image.Through the deep learning model framework,it overcomes the shortcomings of traditional image processing algorithms,achieves accurate and efficient classification of hazardous waste,and can better identify multiple types of hazardous waste.Experiment shows that the proposed method has an accuracy rate of 97.47% on the harmful-waste data set,the model training time is shorter,and its performance is better than other algorithm models.Using deep learning methods to deploy automated garbage classification models is of great significance to environmental protection.
Point Cloud Feature Line Extraction Algorithm Based on PCPNET
YU Meng-juan, NIE Jian-hui
Computer Science. 2022, 49 (11A): 210800017-6.  doi:10.11896/jsjkx.210800017
Abstract PDF(4104KB) ( 1088 )   
References | Related Articles | Metrics
Feature line extraction is the basic operation of geometric model processing,which is of great significance to the expression of 3D model.Based on PCPNET,a calculation method of curvature value and principal curvature direction which is robust to noise and non-uniform sampling is proposed,and a feature line extraction algorithm is proposed.The proposed algorithm uses the weighted quadratic curve to fit the local curvature distribution,and realizes the recognition of ridge and valley feature points by determining the distance from the extreme point of the quadratic curve in the direction of maximum principal curvature.Finally,the minimum spanning tree(MST) of the refined potential feature points is established to connect the feature points and complete the feature line extraction.Experimental results show that the proposed algorithm can use PCPNET to accurately estimate the curvature and principal curvature direction information of point cloud,and according to the proposed feature point recognition method,it can overcome the defect that the traditional simple threshold truncation can not extract the feature lines of flat area normally,and finally extract the feature lines from clean point cloud and noise point cloud accurately and completely.
Image Super-resolution Reconstruction Network Based on Dynamic Pyramid and Subspace Attention
HE Peng-hao, YU Ying, XU Chao-yue
Computer Science. 2022, 49 (11A): 210900202-8.  doi:10.11896/jsjkx.210900202
Abstract PDF(4660KB) ( 1037 )   
References | Related Articles | Metrics
Aiming at the problems of excessive model parameters and reconstruction distortion in existing single-image super-reso-lution convolutional neural networks,a lightweight single-image super-resolution network model based on dynamic pyramid structure and subspace attention module is proposed.First,the network body with dynamic multi-scale pyramid feature combination module consists of dynamic convolution and pyramid grouping convolution.Dynamic convolution can adaptively perform different convolution operations for different images,so as to extract different features for different images.Pyramid grouping convolution not only can better extract multi-scale features,but also can effectively reduce the number of parameters of the network model.Finally,a subspace attention module is used at the end of the network model to divide the channel space of images into multiple subspaces and learn different attention maps for each subspace,which not only can better capture the cross-channel rela-ted information of images,but also allows for effective fusion of image feature information of each subspace.Compared with the existing mainstream algorithms,the proposed method not only has a smaller number of model parameters,but also the reconstructed super-resolution images can achieve better performance in terms of visual effects and quantitative analysis.
Lycium Barbarum Pest Retrieval Based on Attention and Visual Semantic Reasoning
HAN Hui-zhen, LIU Li-bo
Computer Science. 2022, 49 (11A): 211200087-6.  doi:10.11896/jsjkx.211200087
Abstract PDF(3119KB) ( 1021 )   
References | Related Articles | Metrics
Aiming at the problem that traditional retrieval model on pest has a single mode,this paper uses a cross-modal retrieval method for 17 kinds of common lycium pests in image and text modal,which integrates attention mechanism and visual semantic reasoning.First,use Faster R-CNN+ResNet101 to realize the attention mechanism to extract local fine-grained information of wolfberry pest images.Then further introduce vision semantic reasoning to build the image region connections and use convolutional network GCN for region relation reasoning to enhance area representation.In addition,global semantic reasoning is performed by enhancing semantic correlation between regions,selecting discriminant features and filtering out unimportant information to capture more key semantic information.Finally,the semantic association between different modalities of lycium barbarum pest image and text is deeply explored through modal interaction.On the self-built lycium barbarum pest dataset,the average accuracy(MAP) is used as the evaluation index to carry out comparative experiment and ablation experiment.Experimental results demonstrate that the averaged MAP of the proposed method in the self-built lycium pest dataset achieves 0.522,compared with the eight mainstream methods,the average MAP of the method improves by 0.048 to 0.244,and it has better retrieval effect.
Noise Event Classification Model Based on Multimodal Attention
WU He-xiang, WANG Zhong-qing, LI Pei-feng
Computer Science. 2022, 49 (11A): 211000161-7.  doi:10.11896/jsjkx.211000161
Abstract PDF(2588KB) ( 984 )   
References | Related Articles | Metrics
Social media is nowadays one of the main channels for people to obtain news and learn about real-time events due to its low cost,easy access and rapid dissemination.Social media provides a variety of modal information including text and images for analyzing specific events,which contains abundant irrelevant events and false information.To this end,this paper combines the text-image pairs to determine whether the text and image provide information related to specific events,so as to find out irrelevant noise events from the sentence-level of the text.Motivated by the observation that the description in the text is often associated with the scene in the corresponding image,this paper proposes a method of combining text and image information to classify events based on attention mechanism,which can effectively attend to the important information in text and image and promote information interaction in different modalities.Experimental results on CrisisMMD show that our model outperforms six strong baselines,and it can effectively fuse features of different modality to obtain a superior joint representation.
Identification Method of Maize Disease Based on Transfer Learning and Model Compression
DENG Peng-fei, GUAN Zheng, WANG Yu-yang, WANG Xue
Computer Science. 2022, 49 (11A): 211200009-6.  doi:10.11896/jsjkx.211200009
Abstract PDF(3510KB) ( 1058 )   
References | Related Articles | Metrics
Traditional image recognition methods have low accuracy in recognizing maize disease images,and convolutional neural networks have a good effect on image recognition.However,the network model has a large amount of calculation and a large amount of parameters,making it difficult to rely on mobile devices with limited computing power to popularize and use in small sample applications.Therefore,this paper aims to improve the accuracy of maize disease images,reduce network parameters and model size,and proposes a convolutional neural network model that combines migration learning and model compression for maize disease recognition.In order to improve the generalization of the model,the data set is enhanced and the structure of convolutional neural network based on transfer learning is constructed.In this paper,the migration recognition of common maize disease images is carried out by using the improved VGG16-Inception network model pre-trained on ImageNet through transfer learning.Experimental results show that the average recognition accuracy of maize disease images using transfer learning is 93.38% on ImageNet data set.After the migration,the model is compressed by channel pruning and knowledge distillation,and the compressed model is used to recognize the maize disease image by transfer learning.Experimental results show that the average recognition accuracy of corn disease image after compression reaches 92.40%,the accuracy is reduced by 0.98%,the model size is compressed from 73.90 MB to 9.45 MB,and the number of parameters is reduced by 87.80%.The proposed method can ensure the recognition accuracy in small sample scenarios and further realize the model lightweight.
Facial Landmark Fast Detection Based on Improved YOLOv4-tiny
FU Bo-wen, LI Chuang-chuang, LIANG Ai-hua
Computer Science. 2022, 49 (11A): 211100290-5.  doi:10.11896/jsjkx.211100290
Abstract PDF(2666KB) ( 1006 )   
References | Related Articles | Metrics
Facial landmark detection is an important part of face recognition,which has been a hot issue in the field of computer vision.In order to meet the needs of efficient and lightweight face recognition,this paper proposes a facial landmark detection algorithm based on improved YOLOv4-tiny.608*608*3 color image is used for model input.The CSPDarknet53-tiny network is adopted to extract the main features of the input image.Then the extracted features are up-sampled and fused.Attention mechanism is added before feature fusion to improve the detection accuracy.The loss function of YOLOv4-tiny target detection is reconstructed,and the loss function of facial landmark is added to realize the location of facial landmark while detecting.The model output includes face marker frame and five key points.Compared with other facial landmark detection algorithms,the proposed algorithm has higher recognition efficiency and lower configuration requirements while ensuring recognition accuracy.Therefore,it can be better deployed on edge devices or mobile devices.
Thymoma CT Image Prediction Method Based on Deep Learning and Improved Extreme Learning Machine Ensemble Learning
XU Kun-cai, FENG Bao, CHEN Ye-hang, LIU Yu, ZHOU Hao-yang, CHEN Xiang-meng
Computer Science. 2022, 49 (11A): 211200097-6.  doi:10.11896/jsjkx.211200097
Abstract PDF(3580KB) ( 1040 )   
References | Related Articles | Metrics
To predict the risk of thymoma patients before operation,a computer-aided analysis method combining deep learning and extreme learning machine ensemble learning is proposed.Firstly,the CT image of thymoma is transformed to different scales by wavelet multi-scale transform,and the wavelet energy map is calculated to improve the richness and diversity of image information.Secondly,the convolution neural network model is trained by wavelet energy map,and the specific depth features related to tasks in wavelet energy map are extracted by convolution kernel.Finally,the differentiated training subsets are trained based on the improved limit learning machine,and ensemble learning is constructed to improve the stability and prediction accuracy of the model.Based on multicenter experiments,the results show that the proposed method has good generalization performance and stability.The AUCs of the three verification sets are 0.833,0.771 and 0.784 respectively.
Heart Sound Segmentation Algorithm Based on TK Energy Operator and Envelope Fusion
ZHANG Xin, SUN Jing, YANG Hong-bo, PAN Jia-hua, GUO Tao, WANG Wei-lian
Computer Science. 2022, 49 (11A): 210900135-6.  doi:10.11896/jsjkx.210900135
Abstract PDF(2964KB) ( 853 )   
References | Related Articles | Metrics
In order to segment heart sounds by component more effectively,a kind of heart sound segmentation algorithm based on Teager-Kaise energy operator(TKEO) and multi-envelope feature fusion is proposed in experiment.Firstly,the PCG signal is denoised by using the multi-scale wavelet soft threshold.Then TKEO operation is carried out.Since TKEO is extremely sensitive to the instantaneous energy change,the envelope peak can be extracted effectively and the TKEO signal can be obtained.Secondly,the normalized Shannon energy envelope and Viola integral envelope are extracted from the TKEO signal.The Pearson correlation coefficient between each envelope and TKEO signal is calculated.And then the fusion envelope is carried out according to the correlation.Next,the interval search method is used to search the peak envelopes.The variance of the search results is compared.Finally,false peaks are eliminated according to the maximum duration of S1 and S2.The proposed algorithm is tested using PhysioNet2016 data set.Experimental results show that an average accuracy of 0.922 is achieved by using this method.It is proved that this algorithm can be used to segment the heart sound signals effectively.It provides a new method for feature extraction and analysis of heart sound signals collected in clinical environment.
Improved FCOS Target Detection Algorithm
CHEN Jin-ling, CHENG Mao-kai, XU Zi-han
Computer Science. 2022, 49 (11A): 210900220-6.  doi:10.11896/jsjkx.210900220
Abstract PDF(3434KB) ( 1007 )   
References | Related Articles | Metrics
An enhanced FCOS object detection algorithm is proposed to address the problems that the classical anchorless frame object detection algorithm FCOS(fully constitutional one-stage object detection) has difficulty in extracting target information,insufficient ability to combine location and content information,and weak performance due to insufficient differentiation between positive and negative sample.The method first adds a deformable convolution module and a global attention module to the ResNet50 feature extraction network to improve the feature information capture capability.Then,the FPN feature pyramid is combined with the deep link layer to form a multi-scale feature fusion module to improve the feature extraction effect.Finally,the adaptive division of positive and negative samples module is added to enhance the accuracy of the test frame to achieve the effect of improving the regression accuracy.In order to test the detection effect of the algorithm,the COCO dataset and VOC dataset are used for experiments.Compared with the original FCOS algorithm,the average accuracy of the proposed algorithm on the two datasets improves by 2.3% and 1.8%,respectively.Among them,there is a significant improvement for the detection of small targets in the COCO dataset.
Dual Template and Asynchronous Update Tracking Method Based on SiameseFC
MA Han-da, YIN Da
Computer Science. 2022, 49 (11A): 211200133-7.  doi:10.11896/jsjkx.211200133
Abstract PDF(3024KB) ( 837 )   
References | Related Articles | Metrics
SiameseFC has the advantages of fast tracking speed and high accuracy,but it still has some defects in complex scenes,and the tracking mode without updating the template will also cause large errors in the scene that changes rapidly.Therefore,this paper proposes a new tracking method,the dual-template asynchronous update based on SiameseFC.Firstly,both the deep and shallow features are extracted from the VGG-16 network,and two sets of corresponding templates are used respectively,the two sets of templates are updated independently and asynchronously to save computing resources.Then,for the update of the template,the initial template,the template used in the previous tracking,and the template extracted from the tracking result of the previous frame are considered at the same time.And it uses an APCE-based judgment mechanism to dynamically allocate the proportions of the three templatets when updating.This algorithm is superior to mainstream algorithms such as SiamRPN in the benchmark results of OTB100,the success rate and accuracy improve by about 4%～5%,and reaches about 44 fps,which is sufficient to meet real-time tracking requirements.
Study on Recognition and Classification of Congenital Heart Disease and Pulmonary Hypertension Based on ECG Signal
HAN Yu-sen, YANG Hong-bo, SUN Jing, PAN Jia-hua, WANG Wei-lian
Computer Science. 2022, 49 (11A): 210900144-8.  doi:10.11896/jsjkx.210900144
Abstract PDF(5169KB) ( 1065 )   
References | Related Articles | Metrics
Pulmonary arterial hypertension(PAH) associated with congenital heart disease has a high clinical morbidity,disability and mortality.Its diagnosis is mainly made by measuring the mean pulmonary arterial pressure by right heart catheterization.This method is invasive and has high operational requirements,and it is inconvenient to be used in screening,so it is of great significance to explore a non-invasive CHD-PAH intelligent auxiliary diagnosis scheme.This paper studies CHD-PAH on the basis of congenital heart disease,starting from the analysis of ECG signal,and modeling and predicting CHD-PAH by means of preprocessing,heart beat segmentation,waveform detection,feature extraction,data expansion,classification and identification.Based on the Christov_segmenter algorithm,the differential threshold and local peak improvement are used to detect QRS waves,P waves and T waves,and finally extract bimodal features based on time and amplitude.In order to fit the best classification model,the support vector machine,random forest and K-neighbor classifiers are used in experiments,and a sparrow search algorithm based on T distribution is designed to improve the support vector machine.A total of 460 1-lead ECG signals with a duration of 20 s are used for training and testing.Experimental results show that the prediction accuracy,specificity and sensitivity of the SVM model optimized by the proposed algorithm can reach 99.76%,99.80% and 99.73%,respectively.
Rail Surface Defect Detection Model Based on Attention Module and Hybrid-supervised Learning
ZHAO Chen-yang, ZHANG Hui, LIAO De, LI Chen
Computer Science. 2022, 49 (11A): 210800241-6.  doi:10.11896/jsjkx.210800241
Abstract PDF(4781KB) ( 1244 )   
References | Related Articles | Metrics
Rail surface defect detection is an important part of ensuring railway safety.By analyzing the necessity of rail surface defect detection and the shortcomings of existing detection methods,a rail surface defect detection model based on attention mo-dule and hybrid-supervised learning is proposed.Aiming at the problem of a large number of parameters and high deployment cost of existing model,an end-to-end rail defect detection model is proposed.The attention module is used to guide the generation of feature clusters,which improves the speed of defect detection and reduces the cost of model deployment.In view of the problems of few abnormal samples and the high cost of labeling in practical applications,the influence of rough labeling and hybrid supervision is studied,and the pixel-level label data is processed to make different areas of the label get different attention and reduce the dependence of model on label.Finally,experiments are carried out on the actual rail datasets.and the results show that the performance of hybrid-supervised learning is equivalent to that of full supervised learning by adding a small amount of pixel-level label samples to image-level label samples,and the classification accuracy of the model reaches 99.7%.
Supervised Similarity Preserving Deep Second-order Hashing
ZHANG Jian-xin, WU Yue, ZHANG Qiang, WEI Xiao-peng
Computer Science. 2022, 49 (11A): 210900021-8.  doi:10.11896/jsjkx.210900021
Abstract PDF(4705KB) ( 942 )   
References | Related Articles | Metrics
Recently,deep hashing technology,with its advantages of high storage efficiency and quick query speed,has been widely investigated in the field of large-scale visual image retrieval.However,the fundamental image features obtained by current deep hashing methods mainly depend on the first-order statistics of deep convolutional features,and they seldom take the global structure into consideration,leading to the limitation of retrieval accuracy to a certain degree.Focusing on the global representation capability and the similarity of intra-class samples,this paper proposes a novel supervised similarity preserving deep second-order hashing (S2PDSoH) based on deep pairwise supervised hashing,gaining effective performance improvement in the image retrieval task.Based on the pair-wise deep hashing model,S2PDSoH first employs covariance estimation based on matrix power normalization method to capture the deep second-order information of sample images,so that hash codes can possess good global presentation ability.Then,to gain more robust hash codes,it further constructs a joint constraint of category supervision and similarity preservation motivated by the idea of multi-loss integration,followed by an alternate iteration optimization algorithm to realize the end-to-end training.Therefore,with the semantic information added to the dual channel deep second-order hashing framework,S2PDSoH establishes a mechanism for common constraints on category supervision and similarity maintenance.In addition,it also introduces a hash-like function to achieve the approximate binarization result of hash codes,which solves the problem of non-convexity and avoids the quantization error in the hash mapping process.Extensive experimental results on three commonly used data sets show the effectiveness of the proposed deep-order hashing method with supervised similarity preservation.
MIF-CNNIF:A Multi-classification Image Data Framework Based on CNN with Intersect Features
WANG Pan-hong, ZHU Chang-ming
Computer Science. 2022, 49 (11A): 210800267-8.  doi:10.11896/jsjkx.210800267
Abstract PDF(3241KB) ( 854 )   
References | Related Articles | Metrics
In recent years,image multi-classification task and deep learning have received increasingly attentions,and multi-classification image data framework based on convolutional neural network(MIF-CNN) has also been widely used.Traditional CNN-based multi-class image data learning generally has a problem that the image processing is complicated,the feature dimensions are large,and the time complexity is high.To solve this problem,this paper proposes a multi-classification image data framework based on CNN with intersect features(MIF-CNNIF).MIF-CNNIF is a framework for performing multi-classification tasks based on intersect features obtained by multiple feature selection algorithms.Through extensive comparative experiments on 10 multi-class image data sets,the results validate the effectiveness of MIF-CNNIF.The contributions of MIF-CNNIF are that,1)it avoids the problem of setting too many parameters with the usage of pre-trained CNN models;2)it keeps features dimension and time cost after comparing with MIF-CNN;3)it has a better average recognition accuracy than MIF-CNN;4)the effectiveness of combined feature algorithms is verified on multi-class image data sets.
Block-based Compressed Sensing of Image Reconstruction Based on Deep Neural Network
PAN Ze-min, QIN Ya-li, ZHENG Huan, WANG Rong-fang, REN Hong-liang
Computer Science. 2022, 49 (11A): 210900118-9.  doi:10.11896/jsjkx.210900118
Abstract PDF(6444KB) ( 999 )   
References | Related Articles | Metrics
Compressed sensing(CS) is a signal processing framework for effectively reconstructing signal from a small number of measurements obtained by linear projections of the signal.It’s an ongoing challenge for the practicality of computational imaging based on CS.The improvement of image reconstruction model is to incorporate more prior constraints under the signal sparse constraint,and the iterative optimization process is complex and time-consuming.Neural networks,as the application models of deep learning,can realize the approximation of any complex function,which provides a new technical route for high-quality and real-time image reconstruction.In this paper,deep neural network(DNN) is used for reconstruction and the block processing is used to reduce the reconstruction time and the number of network nodes,which avoids the complicated algorithm solving process of CS.The DNN model is obtained by training a large number of different types of images,and then the block CS measurement and DNN nonlinear solution are combined jointly to achieve efficient reconstruction.Experimental results show that,compared with six different reconstruction algorithms,the peak signal-to-noise ratio(PSNR) and structure similarity(SSIM) of images are improved in different degrees.Compared with the advanced CS algorithm,not only the reconstruction quality is comparable,but also the time complexity of DNN is greatly reduced and the reconstruction time is less than 3s.When sampling rate is as low as 0.01,the proposed approach can still reconstruct the image clearly while other algorithms fail.When sampling rate is 0.1,compared with the recent residual network method,the maximum(minimum) gain of PSNR is 6.7(1.97) dB,and the maximum(minimum) gain of SSIM is 0.354(0.122).
Study on Quality Evaluation Method of Speech Datasets for Algorithm Model
LI Sun, CAO Feng, LIU Zi-shan
Computer Science. 2022, 49 (11A): 210800246-6.  doi:10.11896/jsjkx.210800246
Abstract PDF(4253KB) ( 1008 )   
References | Related Articles | Metrics
With the maturity of intelligent voice technology and product application,the demand for high-quality voice datasets is increasing.There have been some researchers put effort on the quality evaluation of the structured data,but there are few stan-dards appeared for the unstructured voice dataset.By analyzing the construction principle of speech algorithm model and analyzing the construction demand of voice dataset,a unified quality assessment framework for the voice dataset is presented.The framework proposes to evaluate the dataset in terms of four dimensions,each of which subsumes a set of criteria:breadth coverage,anthology distinction,field depth and accuracy completeness.The criteria that are suitable to evaluate the quality dimensions are presented,each with the definition,measurement method,and the evaluation process for the voice dataset quality measurement.Experimental assessment and analysis results of the voice datasets in the vehicular application field are presented as the reference for evaluating the voice dataset quality,and promoting the construction of the voice dataset.Considering the diversified applicabi-lity,privacy issues,efficiency requirements,automation requirements and other aspects of the construction of voice data sets,the development suggestions for building high-quality voice datasets are proposed.
Structure Preserved Multi-view Subspace Clustering Based on t-SVD
ZHANG Hua-wei, LU Xin-dong, ZHU Xiao-ming, SUN Jun-tao
Computer Science. 2022, 49 (11A): 210800215-6.  doi:10.11896/jsjkx.210800215
Abstract PDF(2521KB) ( 935 )   
References | Related Articles | Metrics
To peruse the manifold structure and correlation among multi-view data for the tensor based subspace clustering algorithms,this paper proposes a novel algorithm named structure preserved multi-view subspace clustering based on t-SVD(t-SVD-SpMSC).For both structures in multi-view data,we employ the graph regularization in which the graph is got adaptively by iteration.To optimize the objective function,we develop an alternative optimization algorithm to solve the final objective function.The accuracy of clustering using t-SVD-SpMSC on three datasets is 100%,91.51%,99.81% respectively,which shows the priority of the proposed method.
Multi-camera Calibration Method Based on Large-scale Scene
LIAO De, ZHANG Hui, ZHAO Chen-yang
Computer Science. 2022, 49 (11A): 211200054-6.  doi:10.11896/jsjkx.211200054
Abstract PDF(3932KB) ( 1126 )   
References | Related Articles | Metrics
In the field of computer vision,multiple camera sensors need to be used to obtain three-dimensional information of the object so as to achieve a series of measurements such as detection,positioning and size estimation of large target objects.In actual applications,in complex environments,there will be a problem of non-overlapping field of view between the cameras,which prevents effective visual measurement.Therefore,in order to solve the calibration problem of multiple cameras,a multi-camera calibration method with constraints is proposed,without changing the mechanical structure.First,install the cameras in a fixed position,and ensure that there is no vibration between the cameras,by establishing a multi-camera optimization mathematical model and the cameras simultaneously collecting the pose parameter relationships of multiple sets of cameras corresponding to the calibration board,and using the SVRG optimization algorithm to achieve Optimize the pose parameters between the camera coordinate systems,and then obtain the pose matrix between multiple cameras.Finally,the coordinate system transformation matrix between the cameras is used to obtain the relative pose parameters between the corresponding targets as the evaluation accuracy index.And combined with the actual large-scale shield machine to carry out simulation experiment and actual test.The results show that this method has strong anti-interference,stable optimization effect,and can achieve millimeter-level accuracy in practical applications.
Improved Triangular Algorithm for Small Field of View Star Sensor
LU Han, LIN Bao-jun, ZHANG Yong-he, DING Guo-peng, WANG Xin-yu
Computer Science. 2022, 49 (11A): 211000223-5.  doi:10.11896/jsjkx.211000223
Abstract PDF(2321KB) ( 975 )   
References | Related Articles | Metrics
In the fields of scientific exploration and satellite high-precision positioning,star sensors need to obtain attitude information in a short time.In order to improve accuracy,it is necessary to increase the angular resolution,reduce the field of view and use a higher magnitude as a reference.The star pattern recognition algorithm is the key to quickly obtain the attitude information for star sensors.When the traditional triangle algorithm faces a situation where the number of stars is small and the proportion of dark stars is larger,the recognition accuracy will quickly drop to 70%~80%,which needs to be improved.Motivated by this,an efficient star pattern recognition algorithm based on triangle angular distance matching is proposed.The proposed method combines the feature of magnitude interval difference and introduces the fourth star verification,so as to reduce redundant matching.Simulation results show that the proposed method improves the recognition rate to 98.4% while the recognition speed is not lower than that of the classical improved triangle algorithm.In the case of introducing position noise and magnitude noise,it still maintains a recognition rate of more than 93%,which has strong robustness.
Multi-label Vehicle Real-time Recognition Algorithm Based on YOLOv3 and Improved VGGNet
GU Xi-long, GONG Ning-sheng, HU Qian-sheng
Computer Science. 2022, 49 (11A): 210600142-7.  doi:10.11896/jsjkx.210600142
Abstract PDF(2663KB) ( 955 )   
References | Related Articles | Metrics
In order to quickly and effectively identify vehicle information in video,this paper combines the advantages of YOLOv3 algorithm and CNN algorithm to design an algorithm that can identify vehicle multi-label information in real time.Firstly,the high recognition speed and accuracy of YOLOv3 are used to realize real-time monitoring and positioning of vehicles in video stream.After obtaining the vehicle location information,the vehicle information is passed into the improved simplified and optimized VGGNet multi-label classification network to identify the vehicle with multiple tags.Finally,the label information is output to the video stream to obtain real-time multi-label recognition of vehicles in video.The training and test data sets in this paper are derived from KITTI data sets and multi-label data sets obtained through Bing Image Search API.Experimental results show that the mAP of the proposed method on KITTI data set reaches 91.27,the average accuracy of multi-label is more than 80%,and the frame rate of video reaches 35fps.It achieves good results in vehicle identification and multi-label classification on the basis of ensuring real-time performance.
Liver CT Images Segmentation Based on Multi-scale Feature Fusion and Dual AttentionMechanism
HUANG Yang-lin, HU Kai, GUO Jian-qiang, Peng ChengKey Laboratory of Intelligent Computing & Information Processing of Ministry of Education, Xiangtan University, Xiangtan, Hunan 411105,China
Computer Science. 2022, 49 (11A): 210800162-9.  doi:10.11896/jsjkx.210800162
Abstract PDF(3598KB) ( 988 )   
References | Related Articles | Metrics
Liver disease is one of the most common diseases in medicine,and accurate segmentation of liver disease is a necessary step to assist liver disease diagnosis and surgical planning.However,liver segmentation is still a challenging task due to the complexity of liver CT images.With the deepening of research,people begin to consider combining high-level semantics with low-level semantics to further enhance the segmentation effect.However,most of previous studies simply use splicing or summation operation to fuse different semantics,resulting in failure to make full use of its complementarity.To solve the above problems,a network(MD-AUNet) based on multi-scale feature fusion and dual attention mechanism is proposed in this paper.Firstly,the hierarchical dual attention mechanism in the hierarchical multi-scale attention down-sampling module(HAM) is used to effectively fuse feature information of different scales and extract feature representations rich in spatial information.Then,the global context of high-level features is obtained through the global attention up-sampling module(GAM) for weighting the low-level feature information,so as to select more accurate spatial information.At the same time,deep supervision strategy is used in network training to learn the hierarchical representation of different decoding layers.Moreover,a concise and effective post-processing method is proposed to refine the coarse segmentation result of MD-AUNet.Experimental results on the liver datasets collected by the hospital(manually annotated by experts) demonstrate that the proposed algorithm is superior to other existing liver segmentation algorithms in subjective visual perception and objective evaluation indicators,and its mean pixel accuracy,mean IoU and Dice are 97.6%,95.4%,and 95.5% respectively.
Detection of Deepfakes Based on Dual-stream Network
LI Ying, BIAN Shan, WANG Chun-tao, HUANG Qiong
Computer Science. 2022, 49 (11A): 220100106-9.  doi:10.11896/jsjkx.220100106
Abstract PDF(3463KB) ( 1283 )   
References | Related Articles | Metrics
Deepfake is a kind of deep network model based on generative adversarial networks(GAN).It uses the source and target faces to generate highly realistic face videos that are difficult to identify.If some malicious person uses this technology to make fake videos and spread rumors on the Internet,it will infringe personal portrait right,cause adverse social impact,or even cause serious judicial disputes.In view of the serious threat brought by deepfake technology,many researchers at home and abroad pay close attention to the study of the deepfake detection technology,and have put forward some effective detection me-thods.The existing detection methods achieve good detection results in high-quality videos,but most videos in daily applications are usually compressed into low-quality versions through social software.However,most of the existing deepfake detection methods have a significant decline in detecting this kind of low-quality videos.Besides,the detection performances of existing methods are still unsatisfying in the case of cross datasets,limiting their real applications.To address this issue,this paper proposes a dual-flow network structure based on Xception model under multiple attention mechanism.The network structure includes an RGB branch using multiple attention mechanism and a frequency-domain branch for capturing low quality video artifacts.Based on our research,it is found that the tiny difference between real images and fake images tends to concentrates in some local area.The RGB branch under the multiple attention mechanism makes the model focus on different regions of the face,so it can get the glo-bal features aggregated by the low-level texture and high-level semantic features under the guidance of the attention map.Combined with the RGB branch,the discrete cosine transform(DCT) is introduced in the frequency domain branch to provide complementary feature representation,which can reflect subtle forgery traces or compression errors.Specifically,the proposed algorithm firstly extracts a large number of face frames from videos by face extractor algorithm,and feeds these face frames into the two-branch network model.The frequency branch decomposes the spectrum of images with three combined filters that provide additional learnable parts.In the RGB branch,the first three layers of the backbone network extract shallow features including texture information etc.Then the attention module makes the model attend to the shallow information from different local areas.The shallow information is then fed to the attention pooling layer to aggregate with the high-level semantic features from the rest layers of the backbone network.Finally,the network merges the feature vectors from both the RGB-branch and the frequency branch to obtain the final discriminant result.The combination of these two branches can significantly improve the detection performance of the model in cross database scenes and low-quality video sets.In order to verify the effectiveness of the proposed network structure,a large number of comparative experiments are conducted on three public datasets,including FaceForensics++,Celeb DF and DFDC.In the low-quality part of FaceForensics++ dataset,the AUC(Area Under the Curve) can reach 0.9271.In the video level,the detection accuracy of low-quality and high-quality videos can reach 93.84% and 99.69%,respectively.Experimental results show that the proposed algorithm outperforms the existing detection algorithms in low-quality video sets as well as in cross dataset scenes.It verifies that the combination of dual-stream features including the RGB branch and the frequency branch can improve the robustness of the detection method,especially in low-quality video sets and in cross dataset scenes.
Single-stage 3D Object Detector in Traffic Environment Based on Point Cloud Data
CHE Ai-bo, ZHANG Hui, LI Chen, WANG Yao-nan
Computer Science. 2022, 49 (11A): 210900079-6.  doi:10.11896/jsjkx.210900079
Abstract PDF(3342KB) ( 932 )   
References | Related Articles | Metrics
Based on CIA-SSD single-stage 3D object detection model,this paper improves the spatial semantic feature fusion method of the model.A multi-channel fusion module based on attention mechanism is used to fuse the two features.The single-stage detection method TFAF-SSD(two-feature attentional fusion single-stage object detector) is proposed.After extracting the sparse features of point cloud by manifold sparse convolution network,the spatial semantic features of the detected objects are extracted by the spatial semantic convolution layer.After fusion,the output features are predicted.Finally,the final detection frame is output by the detection head.At the same time,different from the previous methods,this paper also uses the data enhancement method to enhance the generalization performance of the model to improve the detection accuracy.The proposed method is verified on KITTI 3D open data set,and the test result of vehicle detection in test set is 83.77%,which is 3.49 % higher than the 80.28% of CIA-SSD model.
Continuous Sign Language Recognition Method Based on Improved Transformer
WANG Shuai, ZHANG Shu-jun, YE Kang, GUO Qi
Computer Science. 2022, 49 (11A): 211200198-6.  doi:10.11896/jsjkx.211200198
Abstract PDF(1457KB) ( 1208 )   
References | Related Articles | Metrics
Continuous sign language recognition is a challenging task.Most current models ignore the overall modeling ability of long sequences,resulting in lower accuracy of recognition and translation of longer sign language videos.The unique codec structure of Transformer model can be used for sign language recognition,but its position coding method and multi-head self-attention mechanism still need to be improved.Therefore,this paper proposes a continuous sign language recognition method based on the improved Transformer model.Through multiple multiplexed position codes with parameters,each word vector in the continuous hand sentence is calculated multiple times to accurately grasp the position information between each word,add learnable memory key-value pairs to the attention module to form a persistent memory module,and expand the number of attention heads and embedding dimensions through linear high-dimensional mapping and the like,to maximize the multi-head attention mechanism of the Transformer model,and the overall modeling ability of long sign language sequences,in-depth mining of key information in each frame of the video.The proposed method achieves competitive recognition results on the most authoritative continuous sign language data sets PHOENIX-Weather2014[1] and PHOENIX-Weather2014-T[2].
Non-contact Heart Rate Detection Based on Facial Video
ZENG Zi-lin, HU Zhi-gang, SHANG Peng, WANG Xin-zheng, FU Dong-liao
Computer Science. 2022, 49 (11A): 211000182-6.  doi:10.11896/jsjkx.211000182
Abstract PDF(3134KB) ( 1139 )   
References | Related Articles | Metrics
Non-contact facial heart rate detection based on video is susceptib