Loading...

Table of Content

    13 December 2021, Volume 0 Issue 11
    Approach for Visual Question Answering Based on Equal Attention Graph Networks
    WANG Tian-xing, YUAN Jia-bin, LIU Xin
    2021, 0(11):  1-6. 
    Asbtract ( 148 )   PDF (1288KB) ( 90 )  
    References | Related Articles | Metrics
    Visual question answering is a task that combines computer vision with natural language processing. It needs to understand the scene in the picture, especially the interaction between different target objects. Great progress on visual question answering has been made in recent years, but traditional methods adopt holistic feature representation, which largely ignores the structure of the given image, and cannot effectively locate objects in the scene. Graph networks rely on high-level image representation, which can capture semantic and spatial relationships. However, the former visual question answering methods using graph networks ignored the role of the correspondence between relations and the question in the answering process. According to this, a visual question answering model based on equal attention graph networks named EAGN is proposed. Relationship edges are given the same importance as object nodes through the equal attention mechanism. The combination of these two elements makes the basis for answering the question more sufficient. Experiments show that compared with other related methods, the EAGN model performs well and is more competitive, which also provides a basis for subsequent related research.
    Prediction of Renal Transplantation Rejection Based on SMOTE and RNN
    YANG Xin-yi, HOU Ling-yan, YANG Da-li, CUI Li-yan
    2021, 0(11):  7-11. 
    Asbtract ( 133 )   PDF (685KB) ( 65 )  
    References | Related Articles | Metrics
    Nowadays, kidney transplantation is more and more widely used, and the prediction of rejection becomes more and more important. In order to solve the problems of high dimension, time sequence and sample imbalance in rejection data, recurrent neural network is applied to predict renal transplantation rejection, in this paper, an algorithm combining SMOTE (Synthetic Minimum Over-sampling Technology) with RNN (Recurrent Neural Network) is proposed. The method first processes the data, reduces the imbalance between samples, and solves the problem of insufficient sample size. Then the key parameters are adjusted and optimized according to the learning process of RNN. The experimental results show that this method can effectively improve the accuracy of positive and negative classification. Compared with the traditional Markov time series prediction algorithm, the accuracy is improved by 16.7%. After the traditional RNN training is optimized, the relative error rate is reduced by 5.03%. This method can be used to effectively predict renal transplantation rejection.
    Text Matching Model Based on BERT and Self-attention Mechanism of Image
    SONG Shuang, LU Xin-da
    2021, 0(11):  12-16. 
    Asbtract ( 213 )   PDF (802KB) ( 96 )  
    References | Related Articles | Metrics
    In order to improve the accuracy of text matching, a text matching model based on BERT(Bidirectional Encoder Representations from Transformers) and self-attention mechanism of image is proposed to overcome the limitations of BERT model and MatchPyramid model in text matching. Firstly, a pair of text is encoded into word-level feature vectors by using the BERT model. Secondly, the matching matrix of word to word similarity between two texts is constructed according to the word vector, which is regarded as a single channel image representation. Then the self-attention feature matrix of the matching matrix is generated by the self-attention mechanism of image. Finally, the matching matrix and the self-attention feature matrix are connected into multi-channel to capture the text matching signals in the image by the convolutional neural network. After the matching signal is connected with the coding vector called [CLS] which is yielded by the BERT model, the similarity of the two texts is obtained by inputting the fully connected neural layer. The experimental results show that the model can effectively improve the MAP and MRR metrics compared with BERT model, MatchPyramid model and other text matching models on WikiQA dataset, and the effectiveness of the model is verified.
    Multimodal Bilinear Pooling Method for Fake News Detection
    LI Guo-dong, PENG Dun-lu
    2021, 0(11):  17-21. 
    Asbtract ( 275 )   PDF (662KB) ( 101 )  
    References | Related Articles | Metrics
    The rise of social media has promoted the development of the news industry and made the spread of fake news more convenient. However, diversified news expressions have brought many negative effects, such as news content exaggerating facts, malicious tampering of news text or image content, the construction of fake news facts arousing public opinion, which makes fake news detection a new challenge in the news field. In order to deal with the research of fake news detection work, the news text and image information are combined, the traditional feature fusion method is changed through the multimodal bilinear pooling method, and a fake news detection model based on the new feature fusion method is constructed. The standard data set in the detection field verifies the performance of the model. The experimental results show that the fusion feature of text and image is irreplaceable in the field of fake news detection, and the proposed method can effectively improve the performance of fake news detection.
    Temporal Context Recommendation Model Integrating Attention and DeepFM
    LIU Yi-xin, WANG Jia-wei, LI Zi-li
    2021, 0(11):  22-27. 
    Asbtract ( 351 )   PDF (837KB) ( 95 )  
    References | Related Articles | Metrics
    For many online E-commerce companies, it is very important to predict the possibility of consumers purchasing goods. Because the interaction between users and commodities is usually high-dimensional and sparse, the deep factor decomposer algorithm (DeepFM) combines the factor decomposer algorithm (FM) with the deep neural network (DNN), uses FM to deal with low-order feature combination, uses DNN to deal with high-order feature combination, and combines the two methods in parallel, which solves the problem of high-dimensional sparse. However, it ignores the order of purchase, that is, time context information. Aiming at this defect, this paper proposes a time context recommendation system (DeepAFM) which integrates attention and DeepFM, which makes better use of the time context information of user and commodity interaction. Compared with DeepFM model without time context information, the AUC is increased by 1.84%. The results of the comparison and verification show that DeepAFM model has better performance.
    Research and Practice on Elastic Scaling of Cloud-Native 5G Network
    ZHAO Shu-jun, HUANG Qian
    2021, 0(11):  28-38. 
    Asbtract ( 256 )   PDF (2234KB) ( 88 )  
    References | Related Articles | Metrics
    As the fast development and wide use of cloud technology, in order to better use the advantages and characteristics of the cloud platform, a lot of cloud-native applications continue to emerge. It is difficult how to solve the problem by using the characteristics of the cloud platform, such as using the elastic scaling. The main container orchestration tool of cloud-native environment is Kubernetes, which supports automatic scaling, but there are some problems that need to be optimized and improved according to specific conditions. This article focuses on researching the horizontal automatic scaling of the 5G network function PCF(Policy Control Function) in Kubernetes environment, by collecting custom customized metrics data (CPU usage, memory usage, transaction count, bandwidth usage), doing a prediction of the future load according to some history load data with LSTM, a feasible elastic scaling algorithm is designed, so as to put forward a method which is perceived in advance, elastic and has no effort on business. A lot of tests and statistics verify the feasibility and correctness of the method.
    Deep Bug Triage Model Based on Multi-head Self-attention Mechanism
    WAN Fa-yang, YU Xu, XU Qi-jiang
    2021, 0(11):  39-43. 
    Asbtract ( 168 )   PDF (1168KB) ( 65 )  
    References | Related Articles | Metrics
    At present, bug tracking system realizes the matching of bug and fixer through bug report. However, the previous bug triage model relies too much on the text quality of the bug report, introduces a lot of redundant information in natural language, and ignores the community relationship between the fixers when the meta field of the bug report is used as the label attribute, which makes the model performance worse. Aiming at the above problems, this paper proposes a multi-head self-attention deep bug triage (MSDBT). The text description of the bug report and the fixer sequence generated from meta field are vectorized; the multi-head self-attention mechanism is used to perform parallel attention calculation among the internal input elements. The results of experiments on four open source software projects show that MSDBT has obvious advantages over the previous model in terms of recall index.
    Research and Application of Flexible Workflow Path Change
    LI Qian-shi, WANG Shu-ying, ZENG Wen-qu
    2021, 0(11):  44-49. 
    Asbtract ( 131 )   PDF (1176KB) ( 65 )  
    References | Related Articles | Metrics
    The flexibility of workflow has become an important issue for workflow system to adapt to the changing business requirements. The application of workflow system is divided into modeling stage and running stage. The static path of existing workflow system can’t adapt to the changing needs of the current enterprise business process path. This paper analyzes the path change problem in the running stage to improve the flexibility of workflow by studying the related theoretical methods and technical solutions. In this paper, the existing workflow path backoff algorithm is improved to support parallel multi-step backoff, which solves the dynamic path problem of workflow system. In addition, the current large number of applications of approval and countersignature function are analyzed, and a solution of dynamically adding parallel branches in operation is proposed, which further improves the path flexibility of workflow. Finally, it is verified by an example. The test results show that the system can dynamically adjust its own path according to the changes of business requirements and operating environment, which greatly improves the flexibility of workflow system.
    Block Target Classification in Remote Sensing Image Based on Active Learning
    QU Xiao-yuan, ZHANG Yong-heng
    2021, 0(11):  50-55. 
    Asbtract ( 136 )   PDF (3200KB) ( 60 )  
    References | Related Articles | Metrics
    In traditional machine learning, the accuracy of the model is often determined by the size of the labeled data sample. However, in the actual situation, only a small part of the massive data is usually accurately marked, while most of the data is not marked. If the data is marked one by one by professionals, it will cost a lot of time and economic costs. Active learning is to retrieve the most useful unlabeled data from a large number of unlabeled data sets, hand it over to professionals for labeling, and then train the model with such samples so as to improve the accuracy of the model. This paper designs a target detection method of remote sensing images. Firstly, a deep learning network model is constructed and pretrained by using the labeled data. This process is iterated repeatedly until the accuracy reaches the set threshold. In the experiment, the labeled data account for 14.2%, 21.4% and 28.6% of the total data respectively. The experimental results show that this method of combining active learing with U-Net network can effectively reduce the amount of data labeling so as to achieve the expected effect of the model.
    Detection of Defective Tablets Based on Binocular Vision
    GUO Zhao-ming, ZHOU Qing-hua, ZENG Xiao-wei
    2021, 0(11):  56-60. 
    Asbtract ( 225 )   PDF (838KB) ( 86 )  
    References | Related Articles | Metrics
    In the traditional pill defect detection method based on machine vision, the pill image is mostly shot from the right above the pill by a single camera, which is difficult to detect tablets with fracture between upper and lower surfaces and unqualified thickness effectively, the thickness of this kind of defective tablets may be different from that of standard tablets. For this kind of situation, this paper presents a method of tablet thickness detection based on binocular vision. In this method, the binocular camera is placed directly above the tablet and the image of the tablet is collected. Canny operator is used to extract the edge features of tablets for stereo matching, the parallax of tablets is calculated by using the principle of binocular vision triangulation, and the depth information of the tablet is calculated according to the parallax, the depth difference between standard tablets and tablets with transverse rupture is calculated, the tablets with thickness defects are detected. The precision of the experimental measurement results is up to 0.1 mm, which has a high detection accuracy and meets the requirements of detection accuracy, indicating that the method has good applicability and reliability.
    An Anchor-free Forest Fire Detection Algorithm Incorporating Attention Mechanism
    LU Ya-nuo, CHEN Bing-cai,
    2021, 0(11):  61-66. 
    Asbtract ( 166 )   PDF (5151KB) ( 110 )  
    References | Related Articles | Metrics
    Forest fire and wildfire are major natural disaster problems, and vegetation is severely damaged all over the world every year. In order to improve the accuracy of forest fire prevention and control, aiming at the problems of complex fire background, low accuracy and low efficiency of traditional methods, this paper proposes a forest fire detection algorithm based on CenterNet. As an anchorless method, CenterNet defines a target as a point and locates the centroid of the target by key point estimation, which can effectively avoid the missed detection of small targets. At the same time, based on an efficient deep feature extraction network, ResNet50, it incorporates an ECA module to suppress useless information and increase the feature extraction capability of the model.Experiments conducted on public forest fire datasets show that compared with other arithmetic methods, the forest fire detection algorithm proposed in this paper has a low false detection rate and a recognition accuracy of 92.39% with a F1 value of 0.86, a Recall value of 79.75%, and a FPS of 43.31.The proposed method has a high detection accuracy and achieves real-time detection of forest fires and implementation of accurate rescue.
    Image Classification Based on Double-pooling Feature Weighted Convolutional Neural Network
    ZHANG Lin-peng, WANG Xi-yuan, LI Qiang
    2021, 0(11):  67-71. 
    Asbtract ( 213 )   PDF (3667KB) ( 79 )  
    References | Related Articles | Metrics
    The traditional pooling method will cause the loss of feature information, resulting in insufficient feature information extracted in the convolutional neural network. In order to improve the accuracy of the convolutional neural network in the image classification process and optimize its learning performance, based on the traditional pooling method, this paper proposes a double-pooling feature weighted structure pooling algorithm, using the maximum pooling and average pooling methods to retain more valuable feature information, and the model is optimized by genetic algorithm. By training convolutional neural networks with different pooling methods, the classification accuracy and convergence speed of convolutional neural networks on different data sets are studied. The experiments compare and verify the classification results of convolutional neural networks using several pooling methods on the remote sensing image data set NWPU-RESISC45 and the color image data set Cifar-10. The result analysis shows that the dual-pooling feature weighting structure makes the classification accuracy of the convolutional neural network be greatly improved, and makes the convergence speed of the model  be further improved.
    Integrated Classification Algorithm Based on Clustering and Undersampling
    ZHOU Chuan-hua, ZHU Jun-jie, XU Wen-qian, DENG Jia-jia
    2021, 0(11):  72-76. 
    Asbtract ( 166 )   PDF (651KB) ( 77 )  
    References | Related Articles | Metrics
    Unbalanced data are often found in various application areas, and traditional classifiers tend to focus on the majority class of samples, which results in unsatisfactory sample classification. To address this problem, an integrated classification algorithm (ClusterUndersampling-AdaCost, CU-AdaCost) based on clustering undersampling is proposed. The algorithm derives the sample centre positions of each cluster by calculating the dimensionally weighted Euclidean distance between samples, and selects the majority class samples with strong information features according to the cluster centroid range to form a new training set. The training set is also placed in an integrated algorithm that introduces a cost-sensitive adjustment function, so as to make the model focus more on the minority class. Through comparison experiments on six UCI datasets, the results show that the algorithm has a strong representation of samples drawn in the undersampling process, which can effectively improve the classification performance of the model for minority categories.
    Discrete Traffic Network Design and Algorithm Based on Tradable Travel Credits with Charging and Rewarding Mechanism
    LIU Bing-quan, LIU Yu-jie, LIU Liang
    2021, 0(11):  77-81. 
    Asbtract ( 117 )   PDF (691KB) ( 61 )  
    References | Related Articles | Metrics
    Tradable travel credits scheme is a kind of novel and more equitable congestion pricing method for transportation network design. In this paper, we will research the network design and management problem according to link-based tradable travel credits with charging and rewarding mechanism and link capacity improvement. A new model of network design and management is formulated by integrating the credits charging and discrete network design. The Logit stochastic equilibrium principle is adopted to capture travelers’ route choice behavior. An algorithm of stochastic equilibrium problem is developed under the feasible constraint of link capacity and credits. As the model of transportation network design and management is established as a mathematical programming with fixed-point constraints model, the particle swarm optimization algorithm is presented to solve it. The model and algorithm are then numerical validated by a network example.
    Dynamic Optimal Path Planning Based on Trajectory Big Data
    ZHANG Xiao-fang, FENG Hui-fang
    2021, 0(11):  82-88. 
    Asbtract ( 180 )   PDF (2693KB) ( 66 )  
    References | Related Articles | Metrics
    Based on trajectory big data, combined with urban traffic status and user personalized needs, a dynamic optimal path planning algorithm is proposed based on improved Viterbi algorithm. First, a traffic network model based on directed complex network with multi-weights is constructed by combining the traffic state with the real road network topology. The multi-weight attributes of the transportation network model are assigned by using the comprehensive weighting method based on the coalition of the analytic hierarchy process and the entropy weight method. Then, a new directed weighted complex network model is obtained. Further, the optimal path is solved by the improved Viterbi algorithm. Finally, taking Lanzhou as an example to analyze the optimal path planning, the effectiveness of the urban optimal path planning algorithm is verified by comparing the proposed algorithm with the static planning method. The experimental results show that path planning that combines urban traffic conditions with user preferences is more scientific and reasonable, and can provide decision-making support and reference for drivers and traffic management departments.
    Improved FCM Algorithm Based on Entropy and Neighborhood Constraint
    FENG Jun-qi, ZHANG Zheng-jun, ZHANG Man, YAN Tao
    2021, 0(11):  89-94. 
    Asbtract ( 152 )   PDF (1589KB) ( 74 )  
    References | Related Articles | Metrics
    Aiming at the problems of fuzzy C-means (FCM) clustering algorithm that does not consider the importance of different attributes of samples and neighborhood information, a FCM algorithm based on entropy and neighborhood constraints is proposed. First the entropy value of each attribute of the sample is calculated to give weight to each attribute, the attribute weight is combined to improve the distance measurement function; then the neighborhood membership weight is calculated according to the distance between the neighborhood sample and the center sample, and the neighborhood membership is got by weighting. The membership degree of the neighborhood constrains the objective function, and the iterative process of the degree of membership is modified, finally the purpose of improving the performance of the FCM clustering algorithm is achieved. Theoretical analysis and experimental results on artificial data sets and multiple UCI data sets show that the improved algorithm is superior to the traditional FCM algorithm, PCM algorithm, KFCM algorithm, KPCM algorithm, and DSFCM algorithm in terms of clustering effect and robustness, which shows the effectiveness of this algorithm.
    A Multi-period Anti-collision Tree Search Algorithm Based on ALOHA Partition
    XUE Wei-lian, LI Xue-jiao, CHEN Jie
    2021, 0(11):  95-99. 
    Asbtract ( 105 )   PDF (961KB) ( 55 )  
    References | Related Articles | Metrics
    In the RFID system, tags are used to transmit information. Within the effective function range of the reader, multiple tags respond at the same time and thus collision is inevitable. The collision of tags will reduce the recognition efficiency of the system. In order to improve the efficiency of label recognition, this paper proposes a multi-cycle anti-collision search algorithm based on ALOHA partition and the existing anti-collision algorithms. Firstly, the frame length of the corresponding time slot is divided according to the number of tags to be recognized. Then, the multi-cycle anti-collision search algorithm is used to identify the tags in the time slot where the collision occurs, which can effectively reduce the probability of collision and improve the efficiency of tag recognition. This algorithm can be applied to a large number of label recognition systems, and has certain advantages over these systems. Theoretical analysis and experimental results show that the algorithm can effectively reduce the number of algorithmic time slots and improve the efficiency of label recognition.
    Named Entity Recognition of Medicinal Plant Texts Integrated with Attention Mechanism
    WANG Yun-qian, WANG Yi-song, CHEN Pan-feng, ZOU Long
    2021, 0(11):  100-105. 
    Asbtract ( 212 )   PDF (1043KB) ( 62 )  
    References | Related Articles | Metrics
    Named entity recognition of medicinal plant texts plays an important role in information extraction and knowledge graph construction in the field of traditional Chinese medicine. Aiming at the problem of long sequence semantic sparsity in medicinal plant attribute text, a disease entity recognition method BAC based on attention mechanism of BiLSTM and CRF model is proposed. Firstly, the medicinal plant attribute text is preprocessed and semi-automatic annotation is used to construct the medicinal plant knowledge data set, and the low-dimensional word vector is obtained by pre-training. Then, these vectors are fed into the attention-based BiLSTM network to obtain feature vectors that better represent disease entities. Finally, the optimal tag sequence is obtained by conditional random field (CRF) algorithm. The comparison of experimental results shows that the accuracy of BAC method reaches 93.78%, which is 4.46% higher than BiLSTM-CRF model, it can effectively improve the recognition effect of named entity of disease in medicinal plant attribute text. The model trained by BAC method is used to identify disease named entities from 1680 text sentences, and a total of 1422 disease entities are extracted. By matching with the names of medicinal plants, a total of 4316 triples of the relationship between medicinal plants name and diseases entities are extracted.
    Automatic Classification of Arrhythmia Based on Improved Residual Dense Network
    LI Chuan-dong, QIU Lei, YU Yan
    2021, 0(11):  106-111. 
    Asbtract ( 128 )   PDF (899KB) ( 59 )  
    References | Related Articles | Metrics
    The automatic classification of different types of arrhythmias can provide reliable diagnosis information for doctors and effectively improve the diagnosis efficiency of this kind of diseases. Therefore, this paper proposes an improved residual dense network for automatic classification of arrhythmias. In the model, the improved residual dense block is used as the basic module of residual dense network, and the depthwise separable convolution is used to replace the traditional convolution to effectively extract the features between channels, which reduces the amount of calculation. At the same time, the channel attention mechanism is introduced to realize the feature selection and improve the weight distribution of important features. Based on the public data set provided by 2018 China physiological signal challenge, nine types of arrhythmias are classified, and the Macro F1_score reaches 81.2%, which is better than the mainstream deep learning network model. The experimental results verify the feasibility and advantages of the model, and provide a new method for arrhythmia classification.
    Monocular Visual Distance Measurement System Used at Park for Collaborative Automatic Parking#br#
    ZHANG Cheng-long, ZHANG Yi
    2021, 0(11):  112-117. 
    Asbtract ( 137 )   PDF (1730KB) ( 72 )  
    References | Related Articles | Metrics
    In the process of realizing collaborative automatic parking, the key point is to identify and locate the target vehicle in the global scope of the parking lot. Therefore, this paper proposes a vehicle identification and location method based on monocular vision that is suitable for the parking lot. Firstly, considering the characteristics of the parking lot, a distributed architecture is designed to ensure the flexibility of the perception system. An improved SSD network is proposed to make it run better in embedded devices, and then a monocular measurement model is established. Finally, the detection and ranging algorithm is used to realize the identification and location of vehicles. In order to verify the effectiveness of the monocular measurement algorithm, an experiment is designed and tested in a real parking lot. The experimental results show that the accuracy of the visual ranging meets the requirements of the system.

    Security Monitoring System Based on Channel State Information
    WANG Li, SUN Xuan-chen, JIANG Shao-wei, E Wan-lin, LI Jia-kang, LU Yu-lan
    2021, 0(11):  118-126. 
    Asbtract ( 139 )   PDF (1732KB) ( 80 )  
    References | Related Articles | Metrics
    CSI is an important index reflecting the spectrum characteristics of each subcarrier signal in the MIMO system. By customizing the Intel 5300 network card driver on the Linux platform, real-time CSI data acquisition can be realized through programming under the conventional wireless network communication environment. CSI information is very sensitive to environmental changes, so it can be used to build a non-contact environment sensing system based on CSI. This paper takes channel state information as data carrier, reviews related applications based on CSI in detail, and establishes an intrusion detection system. The system can successfully detect personnel visit events. The system defines a vector 〖WTHX〗v〖WTBX〗*t={σ, MaxPt, MinPt, E} as the intrusion feature and obtains the value of 〖WTHX〗v〖WTBX〗*t as {1×10-3,[1,2],[1,2],2.0×10-4} through experiments. Meanwhile, when the size of sliding window is set to 10 seconds, the system can achieve a detection success rate of 99.07%.