Loading...

Table of Content

    28 October 2019, Volume 0 Issue 10
    Application of Bidirectional Recurrent Neural Network in Speech Recognition
    Gengzang-Cuomao1,2, HUANG He-ming1,2
    2019, 0(10):  1.  doi:10.3969/j.issn.1006-2475.2019.10.001
    Asbtract ( 182 )   PDF (1199KB) ( 112 )  
    References | Related Articles | Metrics
    In order to solve the problem that feed-forward neural network is difficult to process time series data, bidirectional recurrent neural network (BiRNN) is applied in acoustic modeling of automatic speech recognition. Firstly, the Mel frequency cepstrum coefficients are used for feature extraction. Secondly, bidirectional recurrent neural network is used as acoustic model. And finally, the effects of different parameters on system performance are tested. Experimental results on TIMIT dataset show that, compared with convolutional neural network and deep neural network, the recognition rate of the proposed system is improved by 1.3% and 4.0% respectively, which indicates that BiRNN is more suitable for automatic speech recognition.
    CNN Text Classification Based on Topic Model Word Vectors
    NIU Xue-ying
    2019, 0(10):  7.  doi: 10.3969/j.issn.1006-2475.2019.10.002
    Asbtract ( 109 )   PDF (620KB) ( 77 )  
    References | Related Articles | Metrics
    Mining information in Weibo text is of great significance to automatic question and answer, public opinion analysis and other applied research. The text classification study is the basis of text mining. This paper proposes to input simultaneously the text representations of Word2vec and LDA(Latent Dirichlet Allocation) into convolutional neural network model for high-level semantic feature abstraction and classification learning. The input word vectors can represent both the semantic information between the words and the theme of the text. First, We get the word vectors respectively based on the Word2vec model and LDA. Then the word vectors generated by the two models are cascaded to obtain their text matrix representations. Finally, We put the text matrices into the convolutional neural network simultaneously as two channels to classify the texts, and the effectiveness of the method is verified by experiments on Weibo data.
    Air Group Situation Recognition Method Based on GRU-Attention Neural Network
    GOU Xian-tai, WU Nan-fang
    2019, 0(10):  11.  doi:10.3969/j.issn.1006-2475.2019.10.003
    Asbtract ( 177 )   PDF (2403KB) ( 160 )  
    References | Related Articles | Metrics
    In the modern air battlefield, the results of the intentional determination of the enemy air operations group directly affect our mastery of the situation and the decision-making. Therefore, the assessment of the air group situation is an important task of the modern battlefield. The air combat groups usually perform the corresponding intent according to the mission, monitor the relevant process and mine the corresponding features from the acquired data, and then learn and predict through the intelligent method. This paper proposes a recognition method based on GRU-Attention neural network, which inputs the pre-processed behavior event library into the GRU neural network  to mine deep features. The corresponding weight assignment is automatically calculated by Attention mechanism. Finally, the input information is classified by the softmax layer. The experimental results show that the accuracy of the GRU-Attention situation identification method reaches 96.10%, which verifies the accuracy, efficiency and stability of the proposed method. The proposed method has important theoretical and practical significance for enriching the neural network identification method system and improving the assessment accuracy of the air group configuration potential.
    An XML Keyword Query Algorithm Based on Interval Reserved Coding
    WEI Dong-ping, LUO Dan
    2019, 0(10):  17.  doi:10.3969/j.issn.1006-2475.2019.10.004
    Asbtract ( 138 )   PDF (920KB) ( 83 )  
    References | Related Articles | Metrics
    In recent years, with the explosive growth of XML data, research on XML keyword query technology has received increasing attention. Data coding is the basis of keyword query. There are two main ways at present: path-based coding and interval coding. Interval coding can better adapt to the dynamic updating of XML data in queries, and thus has more advantages. This paper studies the keyword query problem based on interval coding and puts forward a new query algorithm. The algorithm firstly establishes an index according to the reserved interval value, and then selects and traverses the index according to the minimum range value, thereby reducing unnecessary comparison and achieving the purpose of improving query efficiency. The study found that the choice of reserved space has a certain impact on query efficiency. This paper designs an Interval Reservation Based on Node (IRBN) based on the node itself, sets the weight for the node, and sets the interval value according to the weight, forming an interval more balanced coding method according to the node’s own allocation. Experiments show that IRBN coding is reasonable and has high query efficiency.
    Quantitative Evaluation Method for Degree of Decentralization of #br# Blockchain Network Based on Entropy Theory
    WU Ke-ke1, PENG Bo2, XIE Hua2, HUANG Zhen2, YAN Xia1
    2019, 0(10):  21.  doi:10.3969/j.issn.1006-2475.2019.10.005
    Asbtract ( 192 )   PDF (905KB) ( 125 )  
    References | Related Articles | Metrics
    The most primary advantage of blockchain networks is to be decentralization. However, despite the widely acknowledged importance of this property, most studies on this topic lack quantification, and none of them performs a measurement on the degree of decentralization they achieved in practice. Entropy is a measure of uncertainty of random variables, that is the measure of randomness of data sets, can be used to measure the degree of decentralization for blockchain networks. In this paper, taking Bitcoin and Ethereum for instances, we propose an entropy method in information theory to quantify the degree of decentralization for them. Using the information entropy, we calculate the degrees of randomness of blocks mined and address balances to quantify the degrees of decentralization for Bitcoin and Ethereum networks, and the results of calculations indicate that Bitcoin’s mining is more approximately 12% decentralized than Ethereum with full samples, and Bitcoin’s wealth is more approximately 9.2% decentralized than Ethereum with 10000 samples. The method proposed in this paper can be used to evaluate the degree of decentralization for any blockchain network.
    Elastic Load Balancing Algorithm for Cloud Computing Service
    WEN Ting-ting, LI Hong-zhe
    2019, 0(10):  28.  doi:10.3969/j.issn.1006-2475.2019.10.006
    Asbtract ( 216 )   PDF (1134KB) ( 131 )  
    References | Related Articles | Metrics
    Cloud Computing mainly provides resource management and isolation for various applications of users through virtualization technology and in the form of virtual machines, but the overload of virtual machines will reduce the performance of these applications, so it is necessary to balance the load through virtual machine migration to prevent server overload. However, previous load balancing schemes are based on deterministic resource demand estimation and workload characteristics to make migration decisions, without considering the sudden nature of resource demand. In this paper, a flexible load balancing algorithm is proposed by tracking and observing the resource requirements of virtual machines, taking full account of the dynamic and unexpected nature of their workloads. This algorithm effectively solves the problem of inaccurate resource demand estimation and stochastic resource demand forecasting, and provides a new solution for load balancing problem with elastic demand characteristics. Finally, compared with the related algorithms, it shows that the proposed algorithm achieves better results.
    Paper Writing Assistant Platform Based on Microservice Architecture
    OUYANG Hong-ji1, YANG Duo2
    2019, 0(10):  34.  doi:10.3969/j.issn.1006-2475.2019.10.007
    Asbtract ( 176 )   PDF (1226KB) ( 151 )  
    References | Related Articles | Metrics
    In order to solve the problems existing in the process of writing academic dissertations of college students, such as the mismatch between the selected topics and their own academic abilities, the complicated and redundant process of reviewing papers, etc, after investigating and analyzing the writing process of academic dissertations of college students, based on the micro-service architecture and using Spring Cloud, Spring Boot and other technologies, an auxiliary platform for writing academic dissertations is designed and developed, including user login and registration, authentication and authorization, data grabbing, topic selection recommendation, paper online annotation, platform operation and other functions. After actual operation, the platform improves the matching degree between the selected topic and their own academic ability through the recommendation function, realizes the online process of paper review, improves the efficiency of academic paper review.
    Aviation Material Prediction Based on Fuzzy Variable Weight Combination Prediction Model
    YANG Yi-lin, LIU Chen-yu, QI Chang-bin
    2019, 0(10):  40.  doi:10.3969/j.issn.1006-2475.2019.10.008
    Asbtract ( 155 )   PDF (483KB) ( 98 )  
    References | Related Articles | Metrics
    Because of the development of aviation material prediction, there are many kinds of prediction models for each kind of aviation material. The forecasting models of different models reflect the development trends of forecasting objects in different directions. Reasonably processing information can predict the demand trend of aviation materials more comprehensively. Fuzzy variable weight combination forecasting model is different from traditional combination forecasting model, it takes weight coefficient as triangular fuzzy number to combine different prediction model results, uses the criterion of minimizing the prediction accuracy index by single forecasting method, integrates processing to get narrower scope, which can help to forecast decision chain. Finally, an example is given to verify the validity and the accuracy of the model, which can help practical deployment.
    Longitudinal Fragments Stitching Method Based on Correlation Matrix
    LIU Ya-wei, WANG Jun-min, LIU Wei
    2019, 0(10):  43.  doi:10.3969/j.issn.1006-2475.2019.10.009
    Asbtract ( 127 )   PDF (1336KB) ( 92 )  
    References | Related Articles | Metrics
    Due to the use of shredders, there is a need for splicing of shredded paper. In order to make the splicing easier and faster, the shredded paper stitching algorithm based on the correlation matrix in linear algebra is used to automatically splicing and restoring the regular shredded paper printed on one side. The algorithm first converts the image information into a pixel matrix. Then, according to the shape characteristics and the continuity of characters, the relativity of the pixel matrix of the shredded paper edge is calculated, and the correlation matrix is obtained. By comparing the correlation magnitudes, the most relevant piece of debris is obtained for splicing. After experimental verification, the algorithm is simple, feasible and accurate.
    Hybrid Flower Pollination Algorithm For Solving Global Optimization Problems
    ZHU Yang-yang
    2019, 0(10):  48.  doi:10.3969/j.issn.1006-2475.2019.10.010
    Asbtract ( 179 )   PDF (823KB) ( 122 )  
    References | Related Articles | Metrics
    Metaheuristic algorithms can be used as an effective tool for finding near-optimal solutions. Therefore, it is necessary to improve metaheuristic algorithms and enhance the algorithm’s performance. This paper introduces an enhanced variant of Flower Pollination Algorithm (FPA),which combines FPA with Extremal Optimization(EO) to form the FPA-EO. The FPA-EO algorithm makes use of the global search capability of FPA and the local search capability of EO, and applies it to 11 benchmark functions to test the new algorithm. At the same time, the algorithm is compared with other four famous optimization algorithms: standard flower pollination algorithm (FPA), bat algorithm (BAT), firefly algorithm (FA), and simulated annealing algorithm(SA). The comprehensive results show that the algorithm can find a more accurate solution than the other four algorithms.
    An Adaptive Fuzzy Joint Points Clustering Algorithm
    WANG Bao-feng, MA Xiao-xuan, LI Jin-xing
    2019, 0(10):  55.  doi:10.3969/j.issn.1006-2475.2019.10.011
    Asbtract ( 151 )   PDF (868KB) ( 77 )  
    References | Related Articles | Metrics
     Fuzzy Joint Points (FJP) divides the cluster number of clusters by the maximum interval descent method. This method of determining the number of clusters is subjective and is not conducive to the application of the algorithms. Aiming at this problem, an adaptive FJP clustering algorithm based on effective neighbor cluster index is proposed. The Kernels-VCN index is used to evaluate the effectiveness of clustering, so as to achieve the optimal determination of the optimal number of clusters. Finally, we verify the feasibility of the proposed algorithm on UCI datasets and artificial datasets.
    A Multi-features Short-term Traffic Flow Prediction Method Based on SDZ-GRU
    LYU Tian
    2019, 0(10):  60.  doi:10.3969/j.issn.1006-2475.2019.10.012
    Asbtract ( 144 )   PDF (966KB) ( 102 )  
    References | Related Articles | Metrics
    This paper proposes a multi-features short-term traffic flow prediction method based on SDZ-GRU(SGMTFP) to solve the problems that the current short-term traffic flow prediction methods have large errors and only rely on time-series data. This method adds a series of auxiliary data, such as time information, to the existing time-series data. Also, Surprisal-Driven Zoneout (SDZ) is applied to Gated Recurrent Unit (GRU) to make it a new RNN unit called SDZ-GRU. Through the rolling nested cross-validation experiment, the SGMTFP method proposed in this paper is 7.68% and 14.55% lower than the conventional GRU in terms of root-mean-square error and mean absolute error respectively. In addition, since the SGMTFP method adds auxiliary features, compared with the case without auxiliary features, the root-mean-square error and mean absolute error decrease by 10.9% and 15.1%, respectively. The experimental results show that this method can effectively reduce the error.
    An Anomaly Detection Method for Network Traffic of Servers #br# in Smart Grid Based on Deep Encoder-Decoder Neural Network
    YANG Yong-jiao, TANG Liang-liang
    2019, 0(10):  66.  doi:10.3969/j.issn.1006-2475.2019.10.013
    Asbtract ( 136 )   PDF (2256KB) ( 123 )  
    References | Related Articles | Metrics
    Traditonal network traffic anomaly detection is usually based on single original characteristic variable to judge the threshold value, or to judge the threshold value after the dimensionality reduction design statistics of multiple related variables.Although this kind of method is simple, it cannot cope with the nonlinear relationship between variables changing with time. In this paper, a deep neural network is designed for network traffic anomaly detection, which can dynamically identify the non-linear relationship between variables. Two layers of attention mechanism are introduced into Encoder-Decoder neural network, which improves the utilization of long-term historical information and realizes accurate estimation of the normal state of the network traffic. Based on the normal behavior of the estimated network traffic, the distribution of the residual error between the measured value and the estimated value is analyzed, the confidence interval is finally obtained and regarded as the control limit to distinguish abnormal behavior.
    Face Recognition Based on Multi-task Joint Discrimination Sparse Representation
    LI Lei, REN Yue-mei
    2019, 0(10):  72.  doi:10.3969/j.issn.1006-2475.2019.10.014
    Asbtract ( 143 )   PDF (598KB) ( 100 )  
    References | Related Articles | Metrics
    In order to solve the problem that the recognition rate is not high due to the influence of attitude, illumination and noise in face recognition, a face recognition method based on multi-task joint discrimination sparse representation is proposed. Firstly, the local binary features of human face are extracted, and an over-complete dictionary learning objective function of joint classification error and representation error is established based on multiple features. Then, using a multi-task joint discriminant dictionary learning method, the multi-task joint discriminant dictionary and the corresponding classifier are learned. The dictionary has good characterization and discriminant ability, so as to improve the face recognition effect. Experimental results show that the proposed method has better recognition performance than other sparse face recognition methods.
    Face Detection Method Based on YOLO2 for Subway Passenger Flow into Station
    ZHOU Hui-juan, ZHANG Qiang, LIU Yu, WANG Xu-yang, LIU Ying
    2019, 0(10):  76.  doi:10.3969/j.issn.1006-2475.2019.10.015
    Asbtract ( 211 )   PDF (8073KB) ( 113 )  
    References | Related Articles | Metrics
    Due to the problems of illumination change, passenger congestion and large noise interference outside the station, the accuracy of face detection technology for subway passenger flow into station is low nowadays. In order to improve the accuracy of face detection, based on the original network structure of YOLO2 lightweight target detection algorithm Tiny YOLO2, this paper firstly compresses feature maps with different number of 1×1 convolution layers, and then adjusts the size of feature maps to a unified size for cascading to obtain high-dimensional feature maps. We reduce the number of convolution kernels in the last few layers of the network, replace the 3×3 convolution layer of original network with 1×1 convolutional layer to get a deeper and narrower face detection network. The improved network has been trained on the Wider Face dataset and the subway inbound passenger flow dataset to obtain the final face detection model. The trained face detection model is loaded to test 300 randomly selected images of passengers outside the station. The test results show that compared with the Tiny YOLO2 original face detection algorithm, the recall rate is increased by 4.2%, and the detection speed of single image is increased by 6.5%. At the same time, it is tested on the FDDB dataset which is widely used for face detection algorithm evaluation. When the number of false detections is 200, the accuracy of face detection is 5% higher than that of Tiny YOLO 2 and 2% higher than that of SSD. Moreover, this algorithm can achieve a good balance between detection speed and accuracy, and has better generalization.
    Facial Expression Recognition Method Based on Improved LeNet-5
    ZHANG Xiao, ZHOU Lian-zhe, ZHANG Lin-lin
    2019, 0(10):  83.  doi:10.3969/j.issn.1006-2475.2019.10.016
    Asbtract ( 210 )   PDF (1183KB) ( 123 )  
    References | Related Articles | Metrics
    Aiming at the problems of the existing facial expression recognition algorithms, such as long time, slow convergence speed and low classification accuracy, the framework and internal structure of LeNet-5 network are optimized and improved, and a facial expression recognition method based on improved LeNet-5 is proposed. In order to extract more diverse features and improve the ability of feature expression, firstly, the number of convolution layer and pooling layer is increased to adjust the internal parameters of the network; secondly, the generalization ability of the network model is improved by batch normalization of convolution layer and full connection layer; finally, the three pooling layers are overlapped and pooled by the combination of maxpool_avgpool_avgpool. Experiments on FER2013 face expression database show that the improved model has higher recognition accuracy than the current algorithm.
    Image Shadow Detection and Removal Method Based on LAB Color Space
    LIANG Yong-zhen1, PAN Bin2, GUO Xiao-ming2, LIANG Yuan2
    2019, 0(10):  88.  doi:10.3969/j.issn.1006-2475.2019.10.017
    Asbtract ( 342 )   PDF (3267KB) ( 169 )  
    References | Related Articles | Metrics
    In order to achieve a single image’s fast shadow removal, this paper proposes an image shadow detection and removal method based on LAB color space. Firstly, we convert the RGB image into the LAB image, and then detect the shadow image by the edge detection. Then, the image matching the average chromaticity value of the shadow area and the shadowless area is obtained, through analyzing, calculating and re-integrating different color channels. Finally, single image shadow is removed by color correction and edge correction. In order to verify the feasibility and effectiveness of the proposed method, the performance indexes, that is, peak signal to noise ratio (PSNR) and structural similarity (SSIM) are used to evaluate the image shadow removal results objectively. And we compare the proposed method with two typical image shading methods. The results show that the performance index of this method is the highest. In particular, the PSNR performance indexes of three groups of experiments are 17.4721, 17.6206, 17.3048, while the SSIM performance indexes are 0.8192, 0.8344, 0.8027. And the image feature information is clear after shadow removing. Overall, good shadowless effect has been achieved that the retained structure information is closer to the real shadowless scene image.
    Congestion Detection Algorithms for Medicine Box Transportation Based on Machine Vision
    ZHANG Rui-xun1, ZHOU Hong-yu2, LUO Sheng-li2
    2019, 0(10):  94.  doi:10.3969/j.issn.1006-2475.2019.10.018
    Asbtract ( 121 )   PDF (2894KB) ( 98 )  
    References | Related Articles | Metrics
    In the production process of the medicine box, the multi-belt conveyor belt is often used for the delivery of the medicine box, and the congested condition of the medicine box is often occur caused by friction and the other reasons during the process of transportation. Therefore, in order to achieve real-time detection of medicine box congestion, this paper carries out the research and design of the congestion detection system on the conveyor belt based on machine vision technology. Through the collection and analysis of features for industrial production images, the cartridge positioning strategy based on edge detection technology and the vacancy recognition strategy based on gray-scale inversion are designed. And then, the vacancy detection of the kit and the conveyor belt under dynamic background is completed. Finally, this paper designs box congestion detection strategy and congestion balancing strategy, plans the kit deployment route, and solves the conveyor congestion problem.
    A Rotor Surface Defect Detection Method Based on Point Cloud
    LI Yu-meng, DUAN Xin-bo, MA Peng-bo
    2019, 0(10):  101.  doi:10.3969/j.issn.1006-2475.2019.10.019
    Asbtract ( 234 )   PDF (2893KB) ( 176 )  
    References | Related Articles | Metrics
    Rotor defects affect the operation of blower, reduce the working performance and bring safety risks for industrial production. Traditional manual detection is time-consuming and laborious, with low detection and marking accuracy, and it is difficult to accurately classify defects. Therefore, this paper obtains point cloud data based on machine vision, preprocesses it, and compares two defect detection methods respectively based on point cloud registration and workpiece features. The experimental results show that the defect detection based on workpiece features can get more accurate results of defect labeling and classification, and provide a new direction for the research of defect detection methods.
    Non-contact Horse Body Measurement Method Based on Machine Vision
    Maierziyan Abudula, FENG Xiang-ping
    2019, 0(10):  108.  doi:10.3969/j.issn.1006-2475.2019.10.020
    Asbtract ( 218 )   PDF (1399KB) ( 159 )  
    References | Related Articles | Metrics
    At present, the method of horse body measurement refers to the direct contact measurement, which not only consumes lots of manpower and financial resources, but also has the complicated operational steps. In this thesis, the methods of measuring the horse body size are summarized, the methods of herding body size measurement in recent years are reviewed, the main contents of horse body size measurement method are elaborated, and the horse body measurement method is prospected. The paper conducts graying, denoising, thresholding cutting and profile extraction for figures of 10 female horses with different ages, then conducts the image corner detection, and identifies all nodes on the margin of outline, and then uses the pixel traversal to identify the key points of the horse body, including scapula point, forefoot point, sternum front edge point, haunch point and highest haunch point. At the same time, Helon-Qin Jiushao’s formula and geometrical relationship method are used to calculate height and length of the horse body. Such a method is used to measure the average relative error of the horse body length as 3.97% and average relative error of the horse body height as 4.45%. This method provides a scientific research basis for scholars to study the horse body size.
    3D Animation Fusion Technology Based on Block Matching in Color Space
    WANG Ying-li1, ZHAI Jian-li2
    2019, 0(10):  112.  doi:10.3969/j.issn.1006-2475.2019.10.021
    Asbtract ( 161 )   PDF (2288KB) ( 122 )  
    References | Related Articles | Metrics
    In order to improve the image quality of 3D animation, it is necessary to deal with the dynamic information fusion of animation image. A dynamic information fusion technology of 3D animation image based on two-dimensional color space matching is proposed. Virtual scene reconstruction technology is used for 3D animation image acquisition and feature projection processing, binary fitting and edge contour detection for 3D animation image. RGB decomposition technology is adopted for color component extraction of 3D animation image. The color template space projection algorithm is used to deal with the block fusion of 3D animation images to improve the feature matching performance of edge pixels of 3D animation images. Based on the color space block fusion information of 3D animation image, the pixel feature is optimized and the correlation coefficient of matching window is calculated, and the dynamic information fusion processing of 3D animation image is realized. The simulation results show that the proposed method can improve the peak signal to noise ratio (PSNR) of the image output and improve the dynamic imaging quality.
    Intrusion Detection Based on Heterogeneous Convolutional Neural Network
    LI He-ting, FENG Ren-jun, CHEN Hai-yan, JING Dong-sheng
    2019, 0(10):  117.  doi:10.3969/j.issn.1006-2475.2019.10.022
    Asbtract ( 144 )   PDF (895KB) ( 166 )  
    References | Related Articles | Metrics
    Network has penetrated into all fields of people’s production and life. However, due to the existence of a large number of illegal intrusions, the network is facing more and more serious security problems. Therefore, detecting intrusion to ensure network security is an urgent problem to be solved. In order to solve this problem, an intrusion detection method based on heterogeneous convolution neural network is proposed. The convolution neural network model of deep learning is used to extract the intrusion data features. Then the optimal model is obtained according to the training data of convolution neural network with two different structures, which can be used to judge the network intrusion. Finally, experiment on KDD 99 verifies the accuracy and accuracy of the method proposed in this paper.
    Network Intrusion Data Clustering Algorithm Based on Krylov Subspace
    ZHANG Su-ning, WANG Yue-juan, WU Shui-ming, JING Dong-sheng
    2019, 0(10):  121.  doi:10.3969/j.issn.1006-2475.2019.10.023
    Asbtract ( 150 )   PDF (1044KB) ( 152 )  
    References | Related Articles | Metrics
    The data in network information security is generally characterized by high dimension and complex scale. Network intrusion detection requires reasonable analysis of network intrusion information to screen out dangerous aggressive behaviors. With the continuous increase of data dimension, traditional distance-based clustering analysis method is no longer applicable. Therefore, it is more and more important to find an effective method to solve high-dimensional data clustering analysis. A high-dimensional data clustering analysis based on the Krylov subspace methods is proposed, which firstly projects high-dimensional data to lower dimensions, implements the data dimension reduction, and then reoccupies genetic K-means algorithm in low dimensional space for data clustering. The method can not only avoid the loss of data attributes, but also improve the efficiency of high-dimensional data clustering analysis. Finally, experiment on KDD Cup 99 verifies the effectiveness and accuracy of the method.