Computer Science

Survey on Mutual Trust Authentication and Secure Communication of Internet of Vehicles

WANG Chun-dong, LUO Wan-wei, MO Xiu-liang, YANG Wen-jun

Computer Science. 2020, 47 (11): 1-9. doi:10.11896/jsjkx.200800024

Abstract

PDF(2062KB) ( 2628 )

References | Related Articles | Metrics

With the rapid development of multi-scenario applications of the Internet of Vehicles and 5G communications,ensuring mutual trust authentication and secure communication between high-speed vehicles has become increasingly important.The identity authentication in the current Internet of Vehicles access scenario and the securityin the process of communicating with Internet of Vehicles have become the two most important lines of defense.First,this paper introduces the research background of existing mutual trust authentication and secure communication in Internet of Vehicles,and points out the principles and technologies used in secure mutual trust authentication and secure communication,including elliptic curve encryption,Hash function,digital signature,blockchain,etc.Then,it classifies protocols,including anonymous access security mutual trust authentication protocol,group access security mutual trust protocol,cross-domain authentication security mutual trust protocol,etc.Due to the broadcast characteristics of the wireless channel,the information exchanged between vehicle nodes may be eavesdropped,forged or replayed.Therefore,the security communication of Internet of Vehicles based on blockchain and 5G-based light-weight car networking security communication are discussed.Then,it analyzes the existing problems and security threats in the mutual trust authentication and secure communication of the existing car networking.Finally,the impact of 5G communication on the safety certification and communication of the Internet of Vehicles is discussed,and the integration of 5G technology will further develop the mutual trust authentication and secure communication of the Internet of Vehicles in the future.At the same time,it also makes certain predictions and prospects for the future key trends of the combined research of the Internet of Vehicles and 5G technology.

Research on Application of Cryptography Technology for Edge Computing Environment

CHENG Qing-feng, LI Yu-ting, LI Xing-hua, JIANG Qi

Computer Science. 2020, 47 (11): 10-18. doi:10.11896/jsjkx.200500003

Abstract

PDF(1623KB) ( 2147 )

References | Related Articles | Metrics

The sharp increase in the number of edge devices has led to an explosive growth in the amount of data.The centralized data processing model,represented by cloud computing model,has been unable to meet the real-time and high-efficiency requirements of data processing due to its storage characteristics and transmission bandwidth limitations.As the amount of data grows,the importance of edge computing is recognized.Edge computing faces huge security challenges in the development process due to the new features of the edge computing model such as lightweight equipment and heterogeneous architecture.As an important means of protecting information security,cryptography is of great significance for dealing with the security challenges of edge computing.Traditional mature and complete cryptography technologies require corresponding adjustments to the characteristics of edge computing,in order to meet the needs of the new architecture.This paper starts with the security challenges that edge computing model faces,analyzes the corresponding cryptographic technical solutions in the data security field and the application security field,and compares existing research schemes to show the advantages of different technologies in edge computing security protection,which provides new ideas for the application of cryptographic technologies for edge computing.

Implicit Authentication Mechanism of Pattern Unlock Based on Over-sampling and One-class Classification for Smartphones

YAO Mu-yan, TAO Dan

Computer Science. 2020, 47 (11): 19-24. doi:10.11896/jsjkx.200600004

Abstract

PDF(2254KB) ( 1099 )

References | Related Articles | Metrics

Nowadays,smartphones are widely used and stored with sensitive information,and the loss of any personal device can cause fatal information compromise.Thus,the people's attention towards data security has been elevated to a higher level.Considering the delicacy of traditional authentications,this paper investigates an implicit authentication mechanism based on over-sampling and one-class classification,for pattern unlock on smartphones.First,a fusion of time,two-dimensional and three-dimensional sensors is employed,to collect user behavioral biometrics comprehensively.Second,in order to ease the impact caused by noise contained in high-dimensional data,a feature screening,which is composed of feature selection and dimensional compression,is designed.Particularly,in view of the existing limitations of the current binary classification schemes,SVM SMOTE is used to over-sample the user behavioral data,and a one-class classification authentication mechanism is delivered to implement classification,of which the learning process is only based on a single-class diminutive training set.A series of experiments have been conducted on actual data,and results show that the proposed scheme,when only relies on a single-class diminutive training set,performs partially better than the traditional binomial KNN classifier which is trained on large-scale data,in terms of accuracy,FAR,FRR and AUC.

LWID:Lightweight Gait Recognition Model Based on WiFi Signals

ZHOU Zhi-yi, SHONG Bing, DUAN Peng-song, CAO Yang-jie

Computer Science. 2020, 47 (11): 25-31. doi:10.11896/jsjkx.200200044

Abstract

PDF(2352KB) ( 1239 )

References | Related Articles | Metrics

As an important research of pervasive computing and human-computer interaction,identity recognition is widely researched.Although traditional WiFi based identification methods have made good progress,they still face challenges such as limi-ted classification ability,high storage cost and long training time.The above problems motivate us to propose a lightweight gait recognition model based on multi-layer neural networks,which is named as LWID(LightWeight Identification).We firstly reconstruct original time series data into graphs to retain characteristic information among different carriers to the maximum extent.Then we design a bionic Balloon mechanism to tailor neurons in network layer.By combining convolution kernels of different size,we extract data features and integrate channel information in the feature map.The proposed method realizes model scale lightweight with higher classification ability.Experimental results show that the model has 98.8% recognition rate in a 50-person dataset.Compared with traditional WiFi based identification model,LWID has stronger classification ability and robustness.Meanwhile,the model is compressed to 6.14% of current computer vision model size with same accuracy.

Optimized Implementation of Office Password Recovery Based on FPGA Cluster

LI Bin, ZHOU Qing-lei, SI Xue-ming, CHEN Xiao-jie

Computer Science. 2020, 47 (11): 32-41. doi:10.11896/jsjkx.200500040

Abstract

PDF(2079KB) ( 1439 )

References | Related Articles | Metrics

Password recovery is the key technology of password back and electronic forensics.While encrypted office documents are widely used,it is of great significance to achieve the effective recovery of office encrypted documents for information security.Password recovery is a computation-intensive task and requires hardware acceleration to implement the recovery process.Traditional CPUs and GPUs are limited by the processor structure,which greatly limits the further increase in password verification speed.In view of this,this paper proposes a password recovery system based on FPGA cluster.Through detailed analysis of the office encryption mechanism,the password recovery process of each version of office is given.Secondly,the core Hash algorithm is optimized with a pipeline structure on FPGA,the AES algorithm is improved by LUT merging operation,and the password generation algorithm is implemented in parallel at high speed.At the same time,the architecture of FPGA is designed with multiple algorithm sub-modules in parallel,which realizes the fast recovery of office password.Finally,the FPGA accelerator card is used to build the cluster,and the dynamic password segmentation strategy is used to fully explore the low-power and high-performance computing features of FPGAs.The experimental results show that the optimized FPGA accelerator card is more than twice the GPU in terms of computing speed and energy efficiency ratio,which has obvious advantages and is very suitable for large-scale deployment in the cloud to shorten the recovery time and retrieve the password.

Impact of Zipf's Law on Password-related Security Protocols

DONG Qi-ying, SHAN Xuan, JIA Chun-fu

Computer Science. 2020, 47 (11): 42-47. doi:10.11896/jsjkx.200500144

Abstract

PDF(1500KB) ( 863 )

References | Related Articles | Metrics

Identity authentication is the first line of defense for the security of networks and information systems,and password is the most common method of identity authentication.Researches usually assume that user-constructed passwords obey uniform distribution.However,recent studies found that passwords obey Zipf's law,which means that most password-related security protocols underestimate the advantage of an attacker and thus fail to achieve the claimed security.In response to the above problem,first of all,Password-Based Signatures (PBS) protocol proposed by Gjøsteen,et al. and Password-Protected Secret Sharing (PPSS) protocol proposed by Jarecki,et al.are taken as typical representatives.Based on the basic assumption that passwords obey Zipf's law,the security proofs of these two protocols are demonstrated to be flawed,and the security is redefined.Furthermore,the improvements to the two protocols are given respectively.In improved PBS protocol,an attacker's advantage is recalculated.By limiting the guess number of an attacker and entrusting a trusted third party to keep the key,the protocol can prevent a malicious attacker from disguising a legitimate user,and can prevent a malicious server from guessing a user's password and for-ging the signature.In improved PPSS protocol,a Honey_List is set on the server side based on honeywords to detect and prevent online password guessing attack.

Analysis of Large-scale Real User Password Data Based on Cracking Algorithms

XIE Zhi-jie, ZHANG Min, LI Zhen-han, WANG Hong-jun

Computer Science. 2020, 47 (11): 48-54. doi:10.11896/jsjkx.200900077

Abstract

PDF(1942KB) ( 1149 )

References | Related Articles | Metrics

Password authentication is the main authentication method nowadays.It is widely used in various fields,such as finance,military and internet.In this paper,password security is studied from the perspective of an attacker.Large-scale real user data is used for statistical analyses of password general characteristics,and for password vulnerability analyses based on Probabilistic Context-Free Grammars (PCFG) password guessing algorithm and TarGuess-I targeted password guessing model.Through the above analyses,it is found in users' passwords that there are vulnerable behaviors that can be easily discovered and exploited by attackers,such as choosing simple structure passwords,generating passwords based on patterns,password containing semantics and passwords containing personal information (i.e.,name and user name).These vulnerable behavior characteristics are summarized to provide a basis for reminding users to avoid setting weak passwords and studying the method of password strength meter.

Conditional Privacy-preserving Authentication Scheme Based on Blockchain for Vehicular Ad Hoc Networks

XIONG Ling, LI Fa-gen, LIU Zhi-cai

Computer Science. 2020, 47 (11): 55-59. doi:10.11896/jsjkx.200500116

Abstract

PDF(1621KB) ( 1203 )

References | Related Articles | Metrics

With the rapid development of network and information techniques,as an important part of automatic driving,the vehicular ad hoc networks are the core module of the future intelligent transportation system.As a result,the security and conditional privacy of the vehicular ad hoc networks (VANET) has become an urgent research hotspot.However,most of the current conditional privacy-preserving authentication schemes for VANET environment suffer from the problem of cross-datacenter authentication.To the best of our knowledge,blockchain technology has lots of advantages like decentralized and unforgeability bringing a promising solution to this problem compared with the traditional cryptography technologies.However,the current message authentication schemes based on blockchain technology for VANET environment cannot provide unlinkability.To address this issue,this paper designs a lightweight conditional privacy-preserving authentication scheme for VANET environment using physically unclonable function and blockchain technology,which can provide message authentication,integrity,identity privacy preserving,unlinkability and traceability.

Efficient Heterogeneous Cross-domain Authentication Scheme Based on Proxy Blind Signature in Cloud Environment

JIANG Ze-tao, XU Juan-juan

Computer Science. 2020, 47 (11): 60-67. doi:10.11896/jsjkx.191100068

Abstract

PDF(2477KB) ( 1004 )

References | Related Articles | Metrics

In order to solve the problem of identity blindness and efficient heterogeneous cross-domain authentication,an efficient heterogeneous cross-domain authentication scheme based on proxy blind signature is proposed.The scheme reconstructs an efficient and secure cross-domain identity authentication model.Combined with the advantages of proxy signature and blind signature,a trusted certification authority CA is introduced in the cloud to give the third party legal agent the trusted agency authority to perform the proxy blind signature operation.This agent not only reduces the communication load of the inter-cloud certification authority CA,realizes the information interaction between the authorized agent blind signer in different domains and the requesting agent blind signer,but also satisfies the blindness of bidirectional entity identity synchronous authentication and the identi-fiability of the proxy blind signature,and improves the authentication security.The results show that based on the mathematical difficulty,the scheme can meet the performance of anti-substitution attack,resist replay attack,man-in-the-middle attack,identity untraceability and so on,and complete the cross-domain identity authentication with high efficiency and security between foreign users.

Efficient Identity-based Authenticated Key Agreement Protocol with Multiple Private Key Generators

QIN Yan-lin, WU Xiao-ping, HU Wei

Computer Science. 2020, 47 (11): 68-72. doi:10.11896/jsjkx.191000008

Abstract

PDF(1308KB) ( 749 )

References | Related Articles | Metrics

An authenticated key agreement protocol can achieve the authentication and key agreement between users in the secure network communications.In most of large scale network applications,there are multiple Private Key Generators,and a higher-level PKG authenticates the identity and generates a private key for lower-level PKG.Most of the existing identity-based authenticated key agreement protocols with multiple PKGs are designed by using bilinear pairing which needs much more computation resource,and they are also not secure enough.To solve the security and efficiency problems of existing protocols with multiple PKGs,a novel identity-based authenticated key agreement protocol with hierarchical PKGs based on Elliptic Curve Cryptosystem is proposed.In this new scheme,PKGs are not independent to each other,and the lower-level PKG is subordinate to the higher-level PKG.Security analysis show that the proposed protocol can overcome the disadvantages of the existing protocols,and meets security properties such as ephemeral secret leakage resistance,forward security and forgery attack resistance.Comparing with the existing protocols,the novel protocol is free from bilinear paring operation,so it can supply more security with lower computational overhead.

Temporal Reasoning Based Hierarchical Session Perception Recommendation Model

LUO Peng-yu, WU Le, LYU Yang, YUAN Kun-ping, HONG Ri-chang

Computer Science. 2020, 47 (11): 73-79. doi:10.11896/jsjkx.200700088

Abstract

PDF(2260KB) ( 891 )

References | Related Articles | Metrics

Session-based recommendation,which aims at predicting the user's next action based on anonymous sessions,becomes a critical task in many online services.The main challenges of this problem are how to model the temporal relationship of user's behaviors within the target session and capture user's interest by the limited interactions.Existing methods model the user's behavior patterns based on the temporal relationship of adjacent items within the target session,and aggregate the item information in the target session into overall session representation as the corresponding user's interest.In order to improve these two processes,a novel Temporal Reasoning Based Hierarchical Session Perception Model (TRHSP) for session-based recommendation is proposed.On the one hand,unlike the previous works which assume adjacent items are related,TRHSP infers the dependency relationship between adjacent items in the target session and learns to handle the user-item interaction sequence with a flexible order,which helps to model user's behavior.On the other hand,TRHSP aggregates the item information of the target session from both the item level and the item feature level,so as to capture user's interest in a more fine-grained manner.In the experiments on two public datasets,the proposed TRHSP achieves the best performance,thus proving the effectiveness of the model.

Huber Loss Based Nonnegative Matrix Factorization Algorithm

WANG Li-xing, CAO Fu-yuan

Computer Science. 2020, 47 (11): 80-87. doi:10.11896/jsjkx.190900144

Abstract

PDF(3357KB) ( 753 )

References | Related Articles | Metrics

Non-negative matrix factorization (NMF) algorithm can find a non-negative and linear matrix representation and retains the essential characteristics of the original data,it has been successfully applied to many fields.The classical NMF algorithm and its variant algorithms mostly use the mean square error function to measure the reconstruction error,which has been shown to be effective in many tasks,but it still faces some difficulties in dealing with noise-containing data.The Huber loss function performs the same penalty for the smaller residual as the mean square error loss function,and the penalty for the larger residual is linearly grown,so the Huber loss function is more robust than the mean square error loss function.It has been proved that the L_2,1 norm sparse regularization term is a feature selection function in the classification and clustering model of machine learning.Therefore,combining the advantages of the two,a non-negative matrix factorization clustering model based on Huber loss function and incorporating L_2,1 norm regularization term is proposed,and an effective optimization procedure based on projected gradient method to update variables is given.Compared with the classical NMF multi-clustering algorithm on multiple sets of datasets,the experimental results show the effectiveness of the proposed algorithm.

Mixed-sampling Method for Imbalanced Data Based on Quantum Evolutionary Algorithm

YANG Hao, CHEN HONG-mei

Computer Science. 2020, 47 (11): 88-94. doi:10.11896/jsjkx.191000102

Abstract

PDF(2630KB) ( 740 )

References | Related Articles | Metrics

The under-sampling and over-sampling are the common methods for solving the classification problem in an imbalanced data.This paper focuses on the overfitting or lose valuable samples problems brought by using a single sampling method.A mixed sampling method,namely MSQEA,based on quantum evolutionary algorithm is proposed.In MSQEA,the majority class samples and minority class samples are firstly encoded separately to form individuals of population in the quantum evolutionary algorithm,and then an appropriate candidate sampling subset is obtained through optimization iterations.After that,the majority samples in candidate subset are removed by under-sampling to avoid the problem of subsequent oversampling method to generate overmuch redundant samples.Then,an oversampling method is used to generate the minority samples.Additionally,in order to effectively evaluate the fitness of quantum individuals,clustering technique is used to cluster the dataset and the effective validation sets for the evaluation of individuals are obtained.Experiments are conducted to evaluate the performance of algorithm MSQEA.The imbalanced data sets are downloaded from KEEL website,and SMO,J48 and NB are used as classifiers to verify the performance of a classifier after data preprocessing by different sampling methods.Experimental results show that the classification performance of MSQEA is better than some state-of-the art sampling methods.

Domain Label Acquisition Method Based on SL-LDA Model

WANG Sheng, ZHANG Yang-sen, ZHANG Wen, JIANG Yu-ru, ZHANG Rui

Computer Science. 2020, 47 (11): 95-100. doi:10.11896/jsjkx.190900012

Abstract

PDF(1635KB) ( 602 )

References | Related Articles | Metrics

The development of science and technology poses new challenges for the management of literature and scholars.In order to solve the problem of automatic management of massive scientific literature and scholars,this paper proposes a domain label acquisition method based on SL-LDA.On the basis of massive scientific literature,the distribution characteristics of scientificliterature data are analyzed,and the SL-LDA theme model is constructed by introducing the word frequency feature of scientific literature.The theme model is used to extract the “theme-phrase” from the scientific literature of the same scholar and get the initial domain keywords.Then the domain system is introduced,the extraction results of the theme model are vector-represented with the system label.After the position feature weighting,the similarity is used for system mapping.Finally,the domain label of the scholar is obtained.Experiment results show that,compared withthe traditional LDA model,the statistical-based TFIDF algorithm and the TextRank algorithm based on network graph,the final label words obtained by SL-LDA model have better effect and higher accuracy with the same amount of literature data,and the F1 value is also raised to 0.572,indicating that the domain label acquisition method based on SL-LDA has good applicability in the academic field.

Social Network Information Recommendation Model Combining Deep Autoencoder and Network Representation Learning

GU Qiu-yang, JU Chun-hua, WU Gong-xing

Computer Science. 2020, 47 (11): 101-112. doi:10.11896/jsjkx.200400120

Abstract

PDF(3414KB) ( 854 )

References | Related Articles | Metrics

In recent years,using deep learning technology and user-trusted information to improve the recommendation system has become one of the hot topics in the academia,but it is still one of the important challenges to build a model for the recommendation system which combines the two.This paper proposes a hybrid model that expands the deep self-decoder and Top-k semantic social network information by constructing a joint optimization function.The model would collect implicit semantic information based on the network representation learning method and perform experiments with multiple real social network datasets toeva-luate the performance of the AE-NRL model (Autoencoder-Network Representation Learning Model) by various methods.The results show that the model proposed in this paper has better performance than the matrix decomposition method in more sparse and larger data sets.Compared with explicit trust links,the implicit and reliable social network information can better identify the trust degree between users.In the network representation learning technology,deep learning models (SDNE and DNGR) are more effective in the AE-NRL model.

Big Data Intelligent Retrieval and Big Data Block Element Intelligence Separation

HAO Xiu-mei, SHI Kai-quan

Computer Science. 2020, 47 (11): 113-121. doi:10.11896/jsjkx.191000071

Abstract

PDF(1731KB) ( 535 )

References | Related Articles | Metrics

By using∨-type big data structure generated by inverse P-sets,some new concepts of ∨-type big data are given,such as big data block,block matrix,block element,block element matrix and data element.Based on these concepts,the reasoning structure of block attribute,the reasoning structure of block matrix,the intelligent separation theorems,the intelligent retrieval theorems of block element and the equivalence class theorems of block and block element are given.An intelligent separation criterion of block element and an intelligent retrieval criterion of block are presented.A block element intelligent separation-block intelligent retrieval algorithm and its algorithm process are given.The application of big data intelligent retrieval-big data block element intelligent separation-acquisition is given.∨-type big data satisfies the logical characteristic of ‘attribute disjunction'.

Study on Dynamic Adaptive Caching Strategy for Streaming Data Processing

WANG Xu-liang, NIE Tie-zheng, TANG Xin-ran, HUANG Ju, LI Di, YAN Ming-sen, LIU Chang

Computer Science. 2020, 47 (11): 122-127. doi:10.11896/jsjkx.190800093

Abstract

PDF(2023KB) ( 886 )

References | Related Articles | Metrics

In current scenarios of the big data processing application,the streaming data processing technique is widely used.Message middleware or message queue is usually applied as the data buffer in streaming data processing.Apache Kafka is often used as the data buffer middleware.The performance of Kafka largely determines the overall performance of the application system.In practical applications,the streaming data generated by upstream data sources is usually unstable,and the static data caching strategy cannot adapt to this variable production environment.In view of this problem,if there is a strategy that can dynamically adjust the data cache according to the upstream traffic changes,the adaptability of the system to environment can be enhanced,the real-time processing of streaming data caching can be realized and the throughput performance can also be improved.In the dynamic caching strategy,a method of monitoring the upstream data traffic is proposed,and the ARIMA model is used to predict the future traffic of data streaming,so as to adjust the settings of streaming data storage in advance.The optimum setting parameter of streaming data cache comes from multi-objective optimization of the experimental results of middleware system performance under various pressures.Comparative experimental results show that,during the peak period of streaming data,the strategy can improve the throughput performance of Apache Kafka by more than 150% while guaranteeing a certain maximum delay,thus the overall performance of the message middleware system can be improved.

Research Advance on 2D Human Pose Estimation

FENG Xiao-yue, SONG Jie

Computer Science. 2020, 47 (11): 128-136. doi:10.11896/jsjkx.200700061

Abstract

PDF(2212KB) ( 2485 )

References | Related Articles | Metrics

Human pose estimation has always been a research hotspot in the field of computer vision.With the continuous improvement of the performance and accuracy of human pose estimation methods,it can be widely used in human-computer interaction,intelligent surveillance and human activity analysis,etc.In this paper,the methods,models and applications of two-dimensional human pose estimation are reviewed and analyzed,and the future research direction is prospected.The introduction of the method is divided into single person and multi-person pose estimation.In terms of the model,it mainly introduces the models based on ResNet,Hourglass and HRNet.In terms of the application,it mainly introduces the application in the field of human-computer interaction and intelligent surveillance.The research prospect is mainly aimed at the expansion of application scenarios.This paper summarizes the research results in recent years and sorts out the possible research directions.

Underwater Terrain Three-dimensional Reconstruction Algorithm Based on Improved Delaunay Triangulation

CHEN Shi-jie, ZHANG Sen-lin, LIU Mei-qin, ZHENG Rong-hao

Computer Science. 2020, 47 (11): 137-141. doi:10.11896/jsjkx.191100051

Abstract

PDF(2064KB) ( 987 )

References | Related Articles | Metrics

In surface reconstruction of underwater terrain,the commonly used method is to project point cloud data to two-dimensional plane,and use Delaunay triangulation algorithm to generate triangular grid,then combine the depth elevation value in order to restore to three-dimensional space.However,the efficiency of this method is low,and the depth elevation value is abandoned during projection,so it's easy to generate long and narrow triangles in three-dimensional space,which is not conducive to the display effect.Based on the point by point insertion method,the location of the insertion point and the local optimization procedure are improved respectively.The specific performance is as follows:a fusion location algorithm is proposed to find the search direction and locate after calculating the triangle vector area,so as to ensure the uniqueness of the location path and improve the efficiency.At the same time,the depth elevation value is introduced to the local optimization procedure.The standard deviation of triangle angle in three-dimensional space is calculated and used as the standard to measure the similarity to normal triangle.The criterion of empty circumscribed circle is replaced to make the gird more uniform in three-dimensional space.The experimental results show that this method is superior to the traditional Delaunay triangulation algorithm in terms of model quality and construction efficiency.

Scene Text Detection Based on Triple Segmentation

LI Huang, WANG Xiao-li, XIANG Xin-guang

Computer Science. 2020, 47 (11): 142-147. doi:10.11896/jsjkx.200800157

Abstract

PDF(1870KB) ( 807 )

References | Related Articles | Metrics

Scene text detection has been developed rapidly with the development of convolutional neural network.However,there still exists some challenges.On the one hand,many detection algorithms use rectangular box as the detection box,which is inaccurate to locate the irregular texts.On the other hand,some methods may get the bounding boxes but fail to separate text instances that lie very close to each other,causing error detection.To solve these two problems,this paper proposes a novel triple segmentation (TS),text instances in image are mapped to score area,kernel area and threshold area,which generate three segmentation maps,the score map and threshold map are used to guide the generation of kernel map.Although kernel map has the information of texts in image,such as location,size and so on,it lacks the threshold information.In order to get a better result,this method uses threshold map to restrict the generation of kernel map.The detection result is based on instance segmentation to get the bounding polygon of text kernel instance,and then make an expansion.This algorithm achieves a precision of 83% on ICDAR2015 dataset,which outperforms the existing methods by more than 1% on F-measure,which proves this method is also effective to detect curve texts.

Tongue Image Analysis in Traditional Chinese Medicine Based on Deep Learning

LI Yuan-tong, LUO Yu-sheng, ZHU Zhen-min

Computer Science. 2020, 47 (11): 148-158. doi:10.11896/jsjkx.191000104

Abstract

PDF(5490KB) ( 2815 )

References | Related Articles | Metrics

The traditional Chinese medicine tongue diagnosis,because of its intuition and easy to be observed,as well as its high clinical value,convenience and practicability,has become one of the important research subjects.At present,the combination of medical image processing technology,artificial intelligence technology and clinical experience of Chinese medicine experts to achieve objectification,quantification and automation of TCM tongue diagnosis is the mainstream of modernization research of TCM tongue diagnosis.In this paper,the key techniques of tongue segmentation and tongue image feature recognition based on migration learning and deep learning are studied.A tongue segmentation method based on region-based single pixel loss function is proposed.It can instruct the training and learning of the model by combining the color correlation and the semantic correlation between neighboring pixels,and the semantic information of target pixel labels.The experiments show that it partly improves the segmentation effect of the model,the MIoU index on the test set reached 96.32%.Then,a classification model of the tongue image geometric features,which combines spatial transformation network and VGG16 model,is proposed to identify and extract the geometric features of tongue image,providing a basis for syndromic inference of tongue image.Considering the orderliness of the geometric features of the data on the two-dimensional plane,the spatial transformation network is used to explicitly learn the spatial invariance in the model.And the convolution part of the VGG16 model is reused,so that the knowledge learned from the tongue segmentation task can be used for parameter transfer learning.Through two sets of comparative experiments,the validity of the spatial transformation network is proved to improve the spatial invariance of the model,and the knowledge of transfer learning is proved to make the model converge faster and more smoothly.At the same time,a classification model of the tongue image texture features,based on the deep texture coding network and VGG16 model,is proposed to recognize and extract the texture features of tongue image,providing a basis for syndromic inference of tongue image.According to the disorder of texture features in two-dimensional plane,a deep texture coding network is used to encode the ordered feature map,obtained by convolution layers,into an orderless texture semantic representation,which can express texture information more effectively.And the deep texture encoding network can enable the whole model to input images of any size,which gets rid of the loss of texture information caused by scaling operations of fixed input size.The validity of the orderless encoding of the deep texture encoding network for texture semantic representation is verified by the comparative analysis of experiments.

Relative Image Quality Assessment Based on CPNet

LI Kai-wen, XU Lin, CHEN Qiang

Computer Science. 2020, 47 (11): 159-167. doi:10.11896/jsjkx.190900052

Abstract

PDF(3229KB) ( 728 )

References | Related Articles | Metrics

For two images with different quality,the human visual system (HVS) can easily distinguish their quality difference.Thus,it is more accurate to judge the relative quality of two images by simulating HVS than to give the absolute quality score of images.A CPNet (Compare-net) model for evaluating the relative quality between images is proposed in this paper.It is a score-independent algorithm that uses the form of image combination to solve the limitation of data volume.Compared with the absolute quality score label,the proposed relative quality label and relative quality order label have a broader application scenario than the absolute quality score label and are more convenient and accurate to obtain.Firstly,by analyzing the influence of convolutional neural network structure related parameters on network performance,a reasonable network infrastructure is constructed.Secondly,the quality difference characteristics of two images are obtained by the methods of two-channel input network and the feature differentiation,and the classification learning is completed by combining the relative quality labels of the image pairs.Finally,experiments on public database show that the accuracy of the proposed algorithm is better than that of other algorithms.CPNet achieved the best accuracy of 0.971 and 0.947 in the same reference image experiment,and also achieved a very competitive accuracy in different reference image experiments,0.926 and 0.860 respectively.In addition,a three-channel network is designed and experiments are carried out to explore the possibility of extending the proposed algorithm to multiple channels.

Category-specific Image Denoising Algorithm Based on Grid Search

CAO Yi-qin, XIE Shu-hui

Computer Science. 2020, 47 (11): 168-173. doi:10.11896/jsjkx.190900004

Abstract

PDF(2453KB) ( 697 )

References | Related Articles | Metrics

Aiming at the problems of partial region texture loss and time-consuming in similar block search of the category-speci-fic image denoising algorithm,a new denoising algorithm for category-specific image based on grid search is proposed.Firstly,the SSIM is used to select candidate data set similar to the noise image in category-specific data sets.In order to speed up the search of similar blocks,the candidate image set is traversed by a coarse-scale grid search box,and the kNN algorithm is used to find the candidate block in the grid that is close to the noise block.Next,in order to find a candidate block that is closer to the noise block,a fine-scale search box is constructed according to the central position of the candidate block,and the fine-scale search box is traversed to screen the similar block with the closest Euclidean distance between the candidate block and the noise block.Finally,the similar block and the residual component in the regularization of global sparse structure are combined to recover the latent image of the noise image.Experimental results show that the grid search strategy can speed up the selection of similar block,and the residual component can not only remove the image noise,but also better preserve the information at the edge of the image.

Semantic Segmentation Transfer Algorithm Based on Atrous Convolution Discriminator

YANG Pei-jian, WU Xiao-fu, ZHANG Suo-fei, ZHOU Quan

Computer Science. 2020, 47 (11): 174-178. doi:10.11896/jsjkx.191100014

Abstract

PDF(2846KB) ( 628 )

References | Related Articles | Metrics

Supervised semantic segmentation with convolutional neural networks has made great progress in recent years.Since the pix-level labeling required by supervised sematic segmentation is tedious and labor intensive,one way that becomes recently prevalent is to collect photo-realistic synthetic data from video games,where pixel-level annotation can be automatically generated.Despite this,the intrinsic domain difference between synthetic and real images usually causes a significant performance drop when applying the learned model to real world scenarios.To solve this problem,we propose a novel domain adaptive semantic segmentation method.It firstly performs image style conversion over the source domain for reducing the domain difference.Then,the generative adversarial network is employed for feature alignment between source and target domains.In particular,we propose to use the atrous convolution for constructing the powerful discriminator network with the enlarged field of view.Extensive experiments show that the proposed algorithm can achieve 4.5% mIoU improvement on the GTA5 dataset and 2.6% on the SYNTHIA dataset,compared with the classic AdaptSegNet algorithm.

Salient Object Detection Based on Multi-scale Deconvolution Deep Learning

WEN Jing, LI Yu-meng

Computer Science. 2020, 47 (11): 179-185. doi:10.11896/jsjkx.190900008

Abstract

PDF(3092KB) ( 891 )

References | Related Articles | Metrics

Saliency detection aims to highlight the regional objects that people pay attention to subjectively in images.However,the traditional methods mainly distinguish the objects against the background under single resolution,so it's a hard to obtain the local detailed information under various scale.In this paper,we proposed a multi-scale convolution-combined-deconvolution network model.More specifically,we applied the deconvolution on the feature layers as well as their contract features,so that more multi-scale parameters could be maintained;then the fusion of the deconvolution offsets were combined with global information to get the salient result.The experimental results show that with many uncertainty factors in the complex background,compared with traditional methods,the proposed method could get a satisfactory salient detection,Compared with the latest deep learning methods,there can be relatively clear and accurate areas,which reduces the loss of information to some extent and restores more details,at the same time,the runtime of our method has been accelerated due to the design of the independence between the deconvolution layers.

Low-level CNN Feature Aided Image Instance Segmentation

FAN Wei, LIU Ting, HUANG Rui, GUO Qing, ZHANG Bao

Computer Science. 2020, 47 (11): 186-191. doi:10.11896/jsjkx.191200063

Abstract

PDF(2068KB) ( 565 )

References | Related Articles | Metrics

The popular instance segmentation network,Mask R-CNN,has rough target segmentation boundaries and segmentation contours when performing instance segmentation,which leads to low segmentation accuracy.To solve this problem,a high-precision instance segmentation method is proposed by introducing the low-level features of the network into the segmentation branch of Mask R-CNN.Specifically,it selects the convolutional features from lower layers of feature extraction network at first.And then,it resizes the features to a fixed scale (1/8 of the input image) by interpolation algorithm to form the low-level features.It concatenates the features of original segmentation branch of Mask R-CNN with the features extracted by RoI Align ope-ration from low-level features for current target.Since low-level features introduce more low-level texture and contour information,it can effectively improve the accuracy of instance segmentation.Compared with Mask R-CNN,the proposed method obtains 1.2% relative average precision (AP) improvement on the COCO2017 dataset by using ResNet-101-FPN as the feature extraction network.Experimental results show that the proposed method is robust and effective when using different feature extraction networks.

Content-aware Image Retargeting Algorithm Based on Explicit SURF Feature Preservation

ZHAO Liang, PENG Hong-jing, DU Zhen-long

Computer Science. 2020, 47 (11): 192-198. doi:10.11896/jsjkx.191000101

Abstract

PDF(3574KB) ( 676 )

References | Related Articles | Metrics

Image retargeting is a digital media processing technique that adjusts the image size to fit the target resolution of any display device aspect ratio.Most of the existing research on image retargeting focuses on the shape preservation of important objects,while the key features of images are sensitive to the human visual system are not fully considered,resulting in lower visual acceptance.Therefore,a new image retargeting algorithm based on explicit SURF feature preservation is proposed.Different from the general mesh deformation technology based on vertex or axis alignment,the mesh deformation technique based on mesh edge is adopted.First,an affine matrix is defined,so that each mesh edge is deformed according to the affine matrix to form a basic mesh edge-based deformation model.Then,the SURF feature region is obtained by SURF feature detection,and the mesh edge range is constrained to the SURF feature region to achieve the feature preservation effect.Thereby,a new mesh deformation mo-del is obtained.In addition,a sparse energy term is set on the basis of the basic mesh deformation model,that is,assigning initial weights to each mesh edge to make the grid lines sparse each other,thereby solving the problem of grid line self-intersection.This weight can also be updated during the iterative solution process if necessary.Finally,an image quality assessment is performed between the proposed method and two existing image retargeting methods.The proposed method can minimize the distortion and produce better visual effects during the image retargeting process.The highest score gain can reach 16.0% and 9.7%,respectively.

Sketch-based Image Retrieval Based on Attention Model

LI Zong-min, LI Si-yuan, LIU Yu-jie, LI Hua

Computer Science. 2020, 47 (11): 199-204. doi:10.11896/jsjkx.190800145

Abstract

PDF(1855KB) ( 790 )

References | Related Articles | Metrics

To solve the problems of the sparse features and the geometric distortion of hand-drawn images in the research field of SBIR (sketch based image retrieval),a new feature extraction method based on attention model is proposed in this paper.The retrieval results can be obtained efficiently and accurately by accurately extracting the semantic features of hand-drawn images.Firstly,convolutional neural network is used as the basic framework for extracting semantic features,and then the supervised training process is carried out.Attention model mechanism is introduced to locate effective semantic features by adding attention block after the last convolution layer of the convolution neural network,and the attention block is composed of spatial attention structure and channel attention structure.Finally,the final feature descriptor is formed by the fusion of semantic features in different layers,to realize high retrieval accuracy.The experimental results on benchmark Flickr15k dataset proves the feasibility and effectiveness of the proposed method.In addition,the proposed attention model can greatly improve the classification accuracy in the task of sketch classification.

Study on Small Target Pedestrian Detection and Ranging Based on Monocular Vision

HUANG Tong-yuang, YANG Xue-jiao, XIANG Guo-hui, CHEN Liao

Computer Science. 2020, 47 (11): 205-211. doi:10.11896/jsjkx.190900078

Abstract

PDF(2820KB) ( 2095 )

References | Related Articles | Metrics

In order to improve the detection and ranging accuracy of long-distance pedestrians under automatic driving scenario,a pedestrian ranging algorithm is proposed based on the target detection of deep learning.Firstly,a redundant graph cutting method is proposed to detect the small target pedestrian by combining with the YOLOV3 model.And then the candidate bounds of all subgraphs are screened many times by the improved bounding box screening algorithm,and finally the pedestrian detection box is obtained.By analyzing the traditional similar triangle ranging algorithm,an improved similar triangle ranging algorithm including pitch and yaw is proposed.Finally,the transverse and longitudinal distances between pedestrian and the current vehicle are mea-sured in real time according to the pedestrian detection results.The experimental results show that,on the validation set BDD 100 K,the mAP of the proposed redundant graph cutting detection model is 6% higher than that of the original YOLOV3 model,and the mAP of small target pedestrian is improved by 3%,and has a better robustness.On the ranging test set collected by on-board camera,the combination of the redundant graph cutting method and the improved ranging algorithm improves the range measurement accuracy by 6.542% compared with the experimental results,which not only realizes long-distance range measurement,but also has higher range measurement accuracy.

LAC-DGLU:Named Entity Recognition Model Based on CNN and Attention Mechanism

ZHAO Feng, HUANG Jian, ZHANG Zhong-jie

Computer Science. 2020, 47 (11): 212-219. doi:10.11896/jsjkx.191000201

Abstract

PDF(2135KB) ( 1018 )

References | Related Articles | Metrics

Text segmentation and word embedding are usually the first step in Chinese named entity recognition,but there is no clear delimiter between Chinese words and words.OOV(out of vocabulary) words such as professional words and uncommon words are severely disturbing the computation of word vectors.Model performance based on word vector embedding is highly susceptible to word segmentation effects.At the same time,most of the existing models use low-speed recurrent neural network which is difficult to meet the requirements of industrial applications.Aiming at the above problems,this paper constructs a named entity recognition model based on attention mechanism and convolutional neural network:LAC-DGLU.To handel the problem of word segmentation,this paper proposes a word embedding algorithm based on Local Attention Convolution (LAC),which alle-viates the dependence of the model on the effect of word segmentation.For the problem of slow calculation speed,this paper uses aconvolutional neural network with gate structure:Dilated Gated Linear Unit (DGLU) to improve the speed of model calculation.The experimental results on several datasets show that the model can increase the F1 value by 0.2% to 2%compared with the existing mainstream model,and the calculation speed can reach more than 1.4to 1.9 times of the existing mainstream model.

Multi-robot Collaborative Obstacle Avoidance Based on Artificial Potential Field Method

CHEN Jun-ling, QIN Xiao-lin, LI Xing-luo, ZHOU Yang-hao, BAO Bin-guo

Computer Science. 2020, 47 (11): 220-225. doi:10.11896/jsjkx.190900026

Abstract

PDF(2013KB) ( 1103 )

References | Related Articles | Metrics

In recent years,with the increasing attention paid to robots,mobile robot technology has gradually become a research hotspot.Robot obstacle avoidance is an important research topic in mobile robotics,and it is one of the basic problems faced by mobile robots.Aiming at the application scenario of multi-robot,the artificial potential field method is optimized based on the full analysis of the existing robot obstacle avoidance algorithms,and the multi-robot obstacle avoidance algorithm MPF and formation obstacle avoidance algorithm AOA are proposed.MPF algorithm optimizes the problem of local minimum point in artificial potential field method,and increases the probability of robot reaching the target point.AOA algorithm combines with the existing formation obstacle avoidance algorithm to improve the efficiency of formation obstacle avoidance.Finally,different experimental environments are designed for MPF and AOA algorithms respectively.Experiment results show that,in different complex obstacle environments,MPF algorithm can guide the robot to the target point effectively and efficiently,while AOA algorithm can provide efficient and stable formation obstacle avoidance under different environmental complexity and number of robots.

Study on Question Processing Algorithms in Visual Question Answering

XU Sheng, ZHU Yong-xin

Computer Science. 2020, 47 (11): 226-230. doi:10.11896/jsjkx.191200015

Abstract

PDF(2099KB) ( 633 )

References | Related Articles | Metrics

At present,there are various researches on the modeling of Visual Question Answering (VQA) tasks,but existing VQA models have a common drawback,i.e. training and reasoning are time-consuming.Research shows that the text processing part of the VQA model is mainly based on LSTM (Long Short Term Memory) networks,and the overall performance of the VQA model is also limited by the LSTM network used for the text processing.Due to the recurrent nature of the LSTM network,the complex data streams in the LSTM network can hardly take advantages of GPU parallel computing to accelerate.Aiming at the above problems,and for the purpose of optimizing the training speed of the model,a new model named SCMP (Simple Conv1d MaxPool1d) is proposed in this paper to replace the LSTM network to deal with incoming natural language questions.The experimental results on the VQA2.0 dataset show that the training speed of the model is 10 times faster than the existing model,and there is no loss for the accuracy of the VQA model.In addition,this paper proposes a novel method for data augmentation of question datasets in VQA2.0 datasets.Experimental results show that data augmentation can improve model prediction performance and accelerate model convergence.The model trained with enhanced data (SCMP) obtains an evaluation score of 63.46% on the validation set,which is better than the existing VQA model.

Knowledge Graph Completion Model Based on Triplet Importance Integration

LI Zhong-wen, DING Ye, HUA Zhong-yun, LI Jun-yi, LIAO Qing

Computer Science. 2020, 47 (11): 231-236. doi:10.11896/jsjkx.200800195

Abstract

PDF(1687KB) ( 1551 )

References | Related Articles | Metrics

Knowledge graph is a popular research area related to artificial intelligence.Knowledge graph completion is the completion of missing entities given head or tail entities and corresponding relations.Translation models (such as TransE,TransH and TransR) are one of the most commonly used completion methods.However,most of the existing completion models ignore the feature of the importance of the triplets in the knowledge graph during the completion process.This paper proposes a novel knowledge graph completion model,ImpTransE,which takes into account the importance feature in triplets,and designs the entity importance ranking method KGNodeRank and the multi-grained relation importance estimation method MG-RIE,to estimate the entity importance and relation importance,respectively.Specifically,the KGNodeRank method estimates the entity node importance ranking by considering both the importance of the associated nodes and the probability that their importance is transmitted,while the MG-RIE method considers multi-order relation importance to provide a reasonable estimate of the overall importance of the relation.ImpTransE takes into account the entity importance and relation importance features of triplets,so that differentle-vels of attention are given to different triplets during the learning process,which improves the learning performance of the ImpTransE model and thus achieves better completion performance.Experimental results show that ImpTransE model has the best completion performance in most of the metrics on the two knowledge graph datasets compared with the five comparison models,and completion performance of different datasets is consistently improved.

Construction and Application of Enterprise Risk Knowledge Graph

CHEN Xiao-jun, XIANG Yang

Computer Science. 2020, 47 (11): 237-243. doi:10.11896/jsjkx.191000015

Abstract

PDF(2860KB) ( 3678 )

References | Related Articles | Metrics

In supporting semantic Web,knowledge graphs have played an important role in many areas such as search engine,intelligent question-answering system,and recommender system.Therefore,they have become a hot topic in the field of artificial intelligence.Knowledge graphs have many advantages in graph display,mining,and computing,which can help enterprises or financial practitioners analyze and make decisions on business scenarios.At present,some companies have applied knowledge graphs in the financial domain,but these knowledge graphs still suffer from incompleteness.And most existing methods only focus on certain aspects when building financial knowledge graphs.Aiming at these problems above,this paper engages a systematic study on the domain knowledge graph and construct an enterprise risk knowledge graph.This paper describes the construction process of domain knowledge graph from the aspects of ontology construction,knowledge extraction,knowledge fusion,and knowledge storage.Based on the enterprise risk knowledge graph,an intelligent question-answering chatbot is developed to realize the retrieval and application of KG.In order to improve the accuracy of the question answering system,a character-based BiLSTM-CRF model for named entity recognition is used.Experimental results show that the character-based BiLSTM-CRF model performs better than the baseline when the sample size is small.

Event Temporal Relation Classification Method Based on Information Interaction Enhancement

ZHOU Xin-yu, LI Pei-feng

Computer Science. 2020, 47 (11): 244-249. doi:10.11896/jsjkx.190900056

Abstract

PDF(1987KB) ( 734 )

References | Related Articles | Metrics

As a branch of information extraction,event temporal relation classification has attracted more and more attention in recent years due to its good auxiliary effect on many natural language processing tasks.At present,the existing neural network approaches lack of consideration for the information interaction between events.To address this issue,this paper proposes a me-thod of event temporal relation classification based on parameter sharing to enhance information interaction between events.This method firstly learns the semantic information and context information of sentences through gated convolutional neural networks (GCNN),and incorporates them into the shortest dependency path sequence as input.Then,it uses Bidirectional long short-term memory network (Bi-LSTM) to encode the input and capture its semantic representation.In addition,it enhances the information interaction between event pairs by parameter sharing.Finally,the obtained semantic representation is input into the fully connec-ted layer,and the softmax function is used for classification prediction.Experimental results on TimeBank-Dense show that the proposed method outperforms most of the existing neural network methods in classification accuracy.

Visual Sentiment Prediction with Visual Semantic Embedding and Attention Mechanism

LAN Yi-lun, MENG Min, WU Ji-gang

Computer Science. 2020, 47 (11): 250-254. doi:10.11896/jsjkx.190800154

Abstract

PDF(1570KB) ( 729 )

References | Related Articles | Metrics

In order to bridge the semantic gap between visual features and sentiments and reduce the impact of sentiment irrelevant regions in the image,this paper presents a novel visual sentiment prediction method by integrating visual semantic embedding and attention mechanism.Firstly,the method employs the auto-encoder to learn joint embedding of image features and semantic features,so as to alleviate the difference between the low-level visual features and the high-level semantic features.Secondly,a set of salient region features are extracted as input to the attention model,in which the correlations between salient regions and joint embedding features can be established to discover sentiment relevant regions.Finally,the sentiment classifier is built on top of these regions for visual sentiment prediction.The experimental results show that,the proposed method significantly improves the classification performance on testing samples and outperforms the state-of-the-art algorithms on visual sentiment analysis.

Deep Learning Hybrid Forecasting Model for Stock Market

ZHANG Yong-an, YAN Bin-bin

Computer Science. 2020, 47 (11): 255-267. doi:10.11896/jsjkx.200500119

Abstract

PDF(3798KB) ( 2518 )

References | Related Articles | Metrics

Without relying on prior knowledge,deep learning can extract high-level abstract features from a large amount of raw data,which is potentially attractive for financial market forecasting.Based on the idea of “Decomposition-Reconstruction-Integration”,this paper proposes a new method of deep learning prediction methodology,and constructs a deep learning hybrid prediction model－CEEMD-LSTM－for forecasting one-step-ahead closing price of stock market.In this model,CEEMD,as a sequence smoothing decomposition module,can decompose the fluctuations or trends of different scales in time series step by step,producing a series of Intrinsic Mode functions (IMFs) with different feature scales.Long and short-term memory network (LSTM),which is suitable for processing time series in deep learning,is adopted to extract advanced and deep features of each IMF and residual term and to predict the return of closing price of the next trading day.Finally,the predicted values of each IMF components and residual term are integrated to obtain the final predicted value.The empirical results of the stock indices from three stock markets of different developed types demonstrates that the proposed model is superior to other benchmark models in two dimensions－predictive error (RMSE,MAE,NMSE) and predicted directional symmetry (DS).

LFNDIT:Learning Boolean Networks from Nondeterministic Interpretation Transitions

HUANG Yi, KONG Shi-ming, WANG Yi-song, ZHANG Ming-yi, MA Xin-qiang

Computer Science. 2020, 47 (11): 268-274. doi:10.11896/jsjkx.200100079

Abstract

PDF(1384KB) ( 516 )

References | Related Articles | Metrics

Boolean network is an important mathematical model for gene regulation.It is an important issue that inferring structure from the interpretation transitions of Boolean network to discover the regulatory relationship between genes.Thus,resear-chers in the field of Boolean networks have been paying attention for a long time.Existing inductive logic program algorithms cannot infer the network structure from a set of nondeterministic state transitions.To this end,LFNDIT is proposed to learn the structure from state transitions under the asynchronous update semantics of Boolean network.First it translates a set of uncertain state transitions into the set of certain state transitions,and then uses the LF1T learning algorithm proposed by Inoue et al to calculate the corresponding normal logic program (Boolean network).The completeness of LFNDIT is proofed.The preliminary experimental results show that the algorithm can effectively calculate the Boolean network structure from the uncertain state transitions,thus it provides a new idea for discovering Boolean network structure.

Semantic Similarity-based Method for Sentiment Classification

MA Xiao-hui, JIA Jun-zhi, ZHOU Xiang-zhen, YAN Jun-ya

Computer Science. 2020, 47 (11): 275-279. doi:10.11896/jsjkx.191000174

Abstract

PDF(1771KB) ( 944 )

References | Related Articles | Metrics

The sentiment lexicon is helpful for sentiment analysis and can be used to classify sentiment by word matching.However,sentiment lexicon has some limitations in terms of vocabulary coverage and domain adaptation.Therefore,this paper proposes a sentiment classification method based on semantic similarity measurement and embedding representation,which calculates the semantic similarity between the text to be classified and the sentiment lexicon,and combines semantic distance and embedding-based features to classify sentiment,so it is helpful to solve the problem of insufficient use of semantic features.In this paper,the performance of sentiment classification is evaluated by the feature vector extraction from word vectors,sentiment lexicon matching and the proposed method.Experimental results show that this method is better than the comparison method.In the corpus of three e-commerce comment tests,the average F1 value of the proposed method reaches 83.46%,an increase of 8.26% compared with the comparison method.Among them,semantic classification extracted by combining word embedding and ECSD(E-Commerce Sentiment Dictionary) has the best effect,with a performance improvement of 9%,indicating that the extracted emotional semantic features can be enriched by combining semantic similarity,and the performance of emotional classification can be effectively improved.

Optimization Method of Electric Vehicles Charging Scheduling Based on Ant Colony

ZHOU Xin-yue, QIAN Li-ping, HUANG Yu-pin, WU Yuan

Computer Science. 2020, 47 (11): 280-285. doi:10.11896/jsjkx.190700129

Abstract

PDF(1860KB) ( 1228 )

References | Related Articles | Metrics

The rapid development of electric vehicles has brought many convenience to people's living travel and logistics transportation,but electric vehicles have the problem of short driving range due to insufficient power.This paper proposes a charging path scheduling optimization algorithm based on the ant colony for electric vehicles to increase the driving range of electric vehicles.In particular,we first adopt the coulomb counting method to calculate the battery remaining amount of the electric vehicle,and calculate the driving energy consumption of the electric vehicle according to the road traffic condition.Then,we establish the corresponding 0-1 integer programming model,and use the path planning algorithm based on ant colony to obtain the optimal charging path for electric vehicles.After the driving path of the electric vehicle is planned,the pheromone on the path is updated,and the optimal solution and the optimal path are obtained through continuous iteration.The simulation results show that compared with other optimization algorithms,the proposed optimization method can effectively reduce the probability of energy consumption in the process of driving,provide an accurate driving path for electric vehicles,and effectively increase the driving range of electric vehicles.

Network Service Tail Latency Analysis Based on M/M/1 Queuing Model

GUO Zi-ting, ZHANG Wen-li, CHEN Ming-yu

Computer Science. 2020, 47 (11): 286-293. doi:10.11896/jsjkx.191200072

Abstract

PDF(2181KB) ( 1965 )

References | Related Articles | Metrics

Studies have shown that in large-scale network service systems,network latency often exhibits a long tail effect,i.e.,a certain percentage of latency are much larger than the average latency of the network service systems.The long tail of latency has caused widespread concern,and can seriously affect user experience and content provider revenue,especially in large interactive network applications that are sensitive to latency.Therefore,the focus of network service system research has experienced a change from focusing on throughput and average latency to the tail latency of the system of interest.However,most of the exi-sting theoretical models focus on average latency,and it is difficult to analyze the tail characteristics of the network service latency.Due to the complexity of existing research,stay time calculations in complex network services lack formal modeling and computational methods.This paper proposes a method of abstracting a complex network into a queuing model.Based on the model,the expressions of stay time distribution in series and parallel scenes are given,and at the same time,the influence of the tail latency on the change of individual sub-components in the model is analyzed.The predicted results of the model analysis are compared with the results of the simulation network,the error does not exceed 2%.

Detecting Group-and-individual Movements of Moving Objects Based on Spatial-Temporal Anchors of Road-network

HAN Jing-yu, XU Meng-jie, ZHU Man

Computer Science. 2020, 47 (11): 294-303. doi:10.11896/jsjkx.191100083

Abstract

PDF(2715KB) ( 612 )

References | Related Articles | Metrics

To detect the movement of moving objects (vehicles) in real-time fashion,every moving object continuously reports its latest positions to the sever and the server keeps the data to answer queries posed by users,which can incur great communication cost and pressing overload on the server.In particular,this method cannot effectively detect the group movements and track the individual objects simultaneously.Therefore,this paper proposes a framework of double-granularity movement detection based on spatial-temporal anchors (DMDSA),which can effectively detect group movement and track individual positions by embedding each moving object into a spatial-temporal grid and reporting its movement pattern to servers whenever the moving object passes the anchor of each grid cell.During the offline stage,movement patterns of each grid cell are mined from historical trajectories by the server.During the running of moving objects,the server detects the group movements using the maximum-likelihood estimation by aggregating the mined movement patterns.Furthermore,independent-anchor policy and sequenced-anchor policy are used to identify the most likely running path,thus tracking each moving object in real-time.The experimental results on synthetic and real data sets demonstrate that DMDSA framework can not only detect the group movement effectively but also track individual object precisely with the great reduction of communication cost.

Sampling Optimization Method for Acoustic Field Reconstruction Based on Genetic Algorithm

XU Feng, SUN Jie, LIU Shi-jie

Computer Science. 2020, 47 (11): 304-309. doi:10.11896/jsjkx.200600167

Abstract

PDF(3558KB) ( 644 )

References | Related Articles | Metrics

The spatial field of ocean acoustic channel parameters can describe the spatial distribution law of underwater acoustic signal propagation in the ocean,which has important guiding significance for underwater acoustic communication location selection,underwater target detection and stealth.For the problem of sampling trajectory optimization in the application of compressive sensing (CS) methods on the acoustic field reconstruction,a sampling optimization method based on a genetic algorithm (GA) is proposed to improve the CS reconstruction accuracy in this paper combining the characteristics of sound field,compressed sensing and the motion characteristics of underwater robot.Firstly,the structure of the CS sampling matrix is analyzed.Then,combining with the kinematic constraint of underwater vehicles,the gene expression and generation method as well as the GA fitness function are defined to support the sampling of underwater vehicles.In simulations,the traveling salesman problem (TSP)-based path from Gaussian random sampling points and the lawnmower sampling path are used for comparison.The results demonstrate that the proposed GA-based sampling method can significantly improve the reconstruction accuracy of acoustic fields.The influences of different sampling rates and different acoustic filed distributions are discussed,which further illustrates the superior performance of the proposed method.

Design of Temporal-spatial Data Processing Algorithm for IoT

XU He, WU Hao, LI Peng

Computer Science. 2020, 47 (11): 310-315. doi:10.11896/jsjkx.200400045

Abstract

PDF(2336KB) ( 653 )

References | Related Articles | Metrics

With the rapid development of Internet of Things (IoT) and 5G technology,there are more and more applications of artificial intelligence based on deep learning,which makes medical imaging,urban security,autonomous driving and other visual fields based on temporal-spatial data become research hotpots in the direction of IoT.At the same time,the video data,picture data,temperature and humidity data and gas concentration data collected by the IoT system also grow rapidly,which eventually makes the processing speed and feedback speed of the IoT system slower and slower.In view of the large amount of temporal-spatial data collected by IoT nodes and the problem of transient anomalies,this paper designs an EPLSN (Exception Processing Long and Short Memory Network) algorithm based on long and short memory network.This paper designs logic structure of the input gate and improves the network model structure,solving the problem of the classification of transient abnormal data and temporal-spatial data,improving the classification accuracy of the IoT temporal-spatial data,and cleaning the abnormal data.According to the characteristics of the temporal-spatial data collected by the IoT sensor,the data is stored in different data blocks.At the same time,a time-series database is used to temporarily store temporal-spatial data,and an IoT search architecture based on temporal-spatial data is proposed.The architecture is suitable for the real-time search system in IoT environment and accele-rates the search speed of the IoT system.

Optimization of Mobile Charging Path of Wireless Rechargeable Sensor Networks Based on Reinforcement Learning

ZHANG Hao, GUAN Xin-jie, BAI Guang-wei

Computer Science. 2020, 47 (11): 316-321. doi:10.11896/jsjkx.200400075

Abstract

PDF(2242KB) ( 692 )

References | Related Articles | Metrics

Wireless sensor networks occupy an important position in environmental perception and target tracking.In order to recharge sensor nodes in time,this paper proposes a low power consumption and high energy efficientcy mobile path charging algorithm based on reinforcement learning.Wireless sensor network uses a mobile charger to charge the sensor nodes.The Q-Lear-ning algorithm and the epsilon-greedy algorithm are combined to complete the charging of all sensor nodes in turn in the shortest path.Existing related researches usually ignore the maximum amount of power that the sensor node itself can withstand,which easily causes the power to exceed the maximum threshold during charging and suspend work,so the charging time of the mobile charger is limited.The result shows that the proposed mobile charging strategy has a higher utility.Compared with the traditional Q-Learning algorithm and the greedy algorithm,the training cycle is greatly reduced and the energy utilization rate is maximized.

Sliding Window-based Network Coding Cooperative Algorithm in MANET

SONG Ying, ZHONG Xian, SUN Bao-lin, GUI Chao

Computer Science. 2020, 47 (11): 322-326. doi:10.11896/jsjkx.191000181

Abstract

PDF(1744KB) ( 681 )

References | Related Articles | Metrics

Mobile Ad Hoc Network (MANET) is limited by the energy of mobile nodes,the bandwidth of communication links,computing and storage capacity.How to improve the communication bandwidth and data throughput of MANETs is still an urgent problem to be solved.Network coding (NC) is a rapidly developing coding technology,which can effectively increase network bandwidth and network traffic in MANET.Sliding-window Network Coding is a variation of NC that is an addition to multipath routing and improves the throughput of MANET.It is proposed a Sliding Window-based Network Coding Cooperative algorithm in MANET(SWNC-CM).The packets at source nodes are transmitted on the cooperative transport mechanism.Then,code nodes encode the received packets and forward the new packets to next node.The destination node decodes the packets received from different paths and recovers the original data.This algorithm mainly focuses on sliding window,random linear network coding and cooperative data transmission.When SWNC-CM algorithm is used,not all data packets need to be coded.Only those data packets in the same window are coded by random network coding method.Gauss elimination method can be used to decode at the receiving node.This method reduces the computational complexity of encoding/decoding.The performance of this SWNC-CM is studied using NS2 and evaluated in terms of the throughput,decoding delay and packet loss probability when a packet is transmitted.The simulations results shows that the multipath diversity achieved with our proposition can significantly improve the network throughput and packet loss probability.

NWI:CSI Based Non-line-of-sight Signal Recognition Method

TIAN Chun-yuan, YU Jiang, CHANG Jun, WANG Yan-shun

Computer Science. 2020, 47 (11): 327-332. doi:10.11896/jsjkx.190900019

Abstract

PDF(2432KB) ( 724 )

References | Related Articles | Metrics

Due to the influence of complex and changeable indoor environment and multipath effects on WiFi signal propagation,a large number of non-line-of-sight paths are generated,which lead to serious signal fading and communication link quality dete-rioration,resulting in low recognition accuracy and complex system implementation.In this paper,a CSI-based NWI (NLOS recognition based on Wavelet Packet Transform Identification) is proposed,which is mainly used for feature extraction of CSI signals,the physical layer information of WiFi,to identify whether there is blocking in the current link.The three-layer wavelet packet is used to decompose the amplitude of CSI signal,the wavelet packet coefficients,wavelet packet energy spectrum,information entropy and logarithmic energy entropy of 8 frequency bands in the third layer are extracted as feature vectors,and the support vector machine is used for classification.Thereby a non-line-of-sight signal is identified.Compared with other methods,the proposed method does not need to pre-process the CSI signal,and the influence of the environment on transmission signals is maximum retained,so as to reflect the indoor environment more realistically.The experimental results show that the recognition accuracy of the proposed method is 96.23% in the dynamic environment and 94.75% in the static environment.It is proved that the CSI signal feature extraction method based on wavelet packet transform can effectively identify non-line-of-sight signals and has high recognition accuracy and universality.