Current Issue

• • 2019, Vol. 42 No. 6 Published date:28 December 2019 Last issue Next issue

  REVIEW

  • Survey on Machine Reading Comprehension
  • WANG Xiao-jie, BAI Zi-wei, LI Ke, YUAN Cai-xia
  • Journal of Beijing University of Posts and Telecommunications. 2019, 42(6): 1-9. DOI:10.13190/j.jbupt.2019-111
  • Abstract ( 1084 )     HTML     　　
  • In order to make clear the recent work of machine reading comprehension (MRC) tasks, they are divided into four types of subtasks with different sources of answers. They are cloze-style, span selection, multi-choice, and answer generation. Previous work on these four different types of subtasks is investigated under a unified framework of encoder-interaction and reasoning-output. Two types of recent developments for the frame are also given. Several challenges on MRC for future work are discussed at the end of the survey.
  • References | Supplementary Material | Related Articles

  • Architecture and Algorithms of Intelligent Dialogue System
  • HUANG Yi, FENG Jun-lan, HU Min, WU Xiao-ting, DU Xiao-yu
  • Journal of Beijing University of Posts and Telecommunications. 2019, 42(6): 10-19. DOI:10.13190/j.jbupt.2019-169
  • Abstract ( 1331 )     HTML     　　
  • Intelligent human-machine dialogue system comprehensively utilizes a number of core technologies in the field of artificial intelligence. In recent years, with the development of basic algorithms such as deep learning and reinforcement learning, the overall structure, algorithm system and application mode of human-machine dialogue system have been changed and improved greatly. In order to sort out and summarize this technology, the architecture of human-machine dialogue system, the core algorithm of human-machine dialogue system, challenges and technical directions in this field are reviewed.
  • References | Supplementary Material | Related Articles

  Papers

  • Welding Defect Detection of X-Ray Images Based on Faster R-CNN Model
  • GUO Wen-ming, LIU Kai, QU Hui-fan
  • Journal of Beijing University of Posts and Telecommunications. 2019, 42(6): 20-28. DOI:10.13190/j.jbupt.2019-097
  • Abstract ( 1703 )     HTML     　　
  • Based on faster region-based convolutional neural networks (R-CNN) model, a classical model in the field of object detection is used to achieve welding defect detection of X-ray images. A great number of X-ray images are collected and sorted out into the dataset, called WDXI, including no-defect type and 7 defect types. Firstly, an improved method can be used to extract the welding area effectively according to the average gray value and the average contrast value per unit area. The adaptive histogram equalization is used for image enhancement and double median blur is used for noise reduction after experimental comparison. Finally, the testing on the pre-trained model is expected and acceptable in the multi-classification problem of welding defects recognition, and not only proves the research value of WDXI, but also contributes towards making an experiment attempt for improving automatic classification and localization of welding defects combined with Faster R-CNN model.
  • References | Supplementary Material | Related Articles

  • Research on Person Re-Identification Based on Deep Learning under Big Data Environment
  • LI Peng, WANG De-yong, SHI Wen-xi, JIANG Zhi-guo
  • Journal of Beijing University of Posts and Telecommunications. 2019, 42(6): 29-34. DOI:10.13190/j.jbupt.2019-124
  • Abstract ( 594 )     HTML     　　
  • Convolutional neural networks produce higher probability of error for person re-identifications. To overcome the shortcomings, a new deep learning method based on capsule networks model for person re-identification was proposed. First, the standard convolutional layers are used to learn discriminative features. Then, several features in different layers are grouped together to form the primary capsules which represent a rich semantic features. After that, a dynamic routing algorithm which is an iterative routing process, is introduced to decide the attribution between primary capsule and digital capsule. To this end, the digital capsule layer is obtained and each capsule can learn to recognize the presence of persons. To highlight the superiorities of the proposed algorithm, extensive experiments are conducted on a series of challenging datasets and show that the algorithm favorably performs against the previous work.
  • References | Supplementary Material | Related Articles

  • Recommended Model for Fusing Multi-Source Heterogeneous Data Based on Deep Learning
  • JI Zhen-yan, SONG Xiao-jun, PI Huai-yu, YANG Chun
  • Journal of Beijing University of Posts and Telecommunications. 2019, 42(6): 35-42. DOI:10.13190/j.jbupt.2019-164
  • Abstract ( 985 )     HTML     　　
  • Considering that Internet information today is diverse and inconsistent in structure, in order to fully utilize the information provided by multi-source heterogeneous data to improve the recommendation accuracy, a hybrid recommendation model based on deep learning was proposed. The model makes a recommendation based on combining ratings, review texts and social network data. The model also adopts deep learning to learn features of reviews and ratings, and then uses social network to constraint sampling. Experiments show that the model is of higher accurate feature representations of users and items.
  • References | Supplementary Material | Related Articles

  • Traffic Distribution Algorithm Based on Multi-Agent Reinforcement Learning
  • CHENG Chao, TENG Jun-jie, ZHAO Yan-ling, SONG Mei
  • Journal of Beijing University of Posts and Telecommunications. 2019, 42(6): 43-48,57. DOI:10.13190/j.jbupt.2019-140
  • Abstract ( 970 )     HTML     　　
  • Most of the researches on traditional traffic engineering strategies focus on constructing and solving mathematical models. To reduce computational complexity,an experience-driven traffic allocation algorithm based on multi-agent reinforcement learning was proposed. It can effectively distribute traffic on pre-calculated paths without solving complex mathematical models and then fully utilize network resources. The algorithm performs centralized training on the software defined networking controller,and can be executed on the access switch or router in a distributed way after the training is completed. Frequent interactions with the controller are avoided at the same time. Experiments show that the algorithm is effective in reducing the end-to-end delay and increasing throughput of the network with respect to the shortest-path and the equal-cost multi-path.
  • References | Supplementary Material | Related Articles

  • Research on Automatic Recognition Algorithm of OBD Port Occupancy State Based on Deep Learning
  • SU Dong, YU Ning-mei
  • Journal of Beijing University of Posts and Telecommunications. 2019, 42(6): 49-57. DOI:10.13190/j.jbupt.2019-180
  • Abstract ( 560 )     HTML     　　
  • Because that optical branching device (OBD) port occupancy state can not be automatically acquired, an improved YOLOv3 algorithm was proposed. Firstly, by adding the fourth upsampling feature map, the detection sensitivity of dense small objects at high resolution is increased. Secondly, for characteristics of the fixed height-width ratio of port, the k-means clustering algorithm is used to re-determine the number of target candidate boxes and the height-width ratio. Thirdly, the soft non-maximum suppression algorithm is proposed to alleviate the missed detection and false detection caused by the proximity and occlusion of the port. Finally, four difficult production scenario in the port occupancy state is detected to verify the performance of the improved YOLOv3 algorithm. Experiments show that the accuracy of the improved YOLOv3 is 90.12%, 5.17% higher than the original YOLOv3. In conclusion, the improved algorithm has higher detection accuracy for port-like objects.
  • References | Supplementary Material | Related Articles

  • Optimization Algorithm of Anchor Node Layout Based on Beetle Antennae Search
  • DENG Zhong-liang, LIU Yan-xu, HU En-wen
  • Journal of Beijing University of Posts and Telecommunications. 2019, 42(6): 58-63. DOI:10.13190/j.jbupt.2019-135
  • Abstract ( 641 )     HTML     　　
  • The anchor node layout scheme determines the performance of localization in the wireless sensor network. In order to solve the problems of high computational cost and limited optimization strategy of existing layout algorithms, an anchor node layout optimization algorithm based on the beetle antennae search was proposed. The algorithm adopts Cramér-Rao lower bound vector as the optimization strategy, and the efficiency coefficient method to optimize the layout strategy and then the algorithm uses beetle antennae search algorithm to achieve rapid deployment of anchor node. The simulations show that the proposed algorithm reduces the low bound of localization performance by 38.79% in 99.74% of the regions, and in the 25-anchor node layout scenario, compared with the genetic algorithm, the low bound of localization performance is almost same, but the search time is reduced by about 64.2%.
  • References | Supplementary Material | Related Articles

  • Tasks Offloading and Resource Scheduling Algorithm Based on Deep Reinforcement Learning in MEC
  • XUE Ning, HUO Ru, ZENG Shi-qing, WANG Shuo, HUANG Tao
  • Journal of Beijing University of Posts and Telecommunications. 2019, 42(6): 64-69,104. DOI:10.13190/j.jbupt.2019-155
  • Abstract ( 699 )     HTML     　　
  • In order to improve the task offloading efficiency in multi-access edge computing (MEC), a joint optimization model for task offloading and heterogeneous resource scheduling was proposed, considering the heterogeneous communication resources and computing resources, jointly minimizing the energy consumption of user equipment, task execution delay, and the payment. A deep reinforcement learning method is adopted in the model to obtain the optimal offloading algorithm. Simulations show that the proposed algorithm improves the comprehensive indexes of equipment energy consumption, delay, and payment by 27.6%, compared to the Banker's algorithm.
  • References | Supplementary Material | Related Articles

  • An ANTLR-Based Feature Extraction and Detection System for Scratch3.0
  • LIU Pai, SUN Yan, REN Wei
  • Journal of Beijing University of Posts and Telecommunications. 2019, 42(6): 70-75. DOI:10.13190/j.jbupt.2019-125
  • Abstract ( 540 )     HTML     　　
  • As a visual programming language for children, Scratch has received wide attention in the programming education. Considering that Scratch has evolved to the latest version 3.0 and its storage structure changes significantly from the previous version, the existing methods cannot be directly applied to project analysis. A new feature extraction and detection system based on linked list data structure and another tool for language recognition (ANTLR) was presented to solve the problem. Experimental results show that the system can effectively extract programming features from the projects and provide feedback to students and teachers. Moreover, its detection performance and stability perform better than the original methods in Scratch2.0.
  • References | Supplementary Material | Related Articles

  • Based on Multiple Probing Tasks Fine-Tuning of Language Models for Text Classification
  • FU Qun-chao, WANG Cong
  • Journal of Beijing University of Posts and Telecommunications. 2019, 42(6): 76-83. DOI:10.13190/j.jbupt.2019-149
  • Abstract ( 631 )     HTML     　　
  • Pre-trained language models are widely used in many natural language processing tasks, but there is no fine-tuning for different tasks. Therefore, for text classification task, the author proposes a method of fine-tuning language model based on probing task, which utilizes the specific linguistic knowledge of probing task training model, and improves the performance of the model in text classification task. Six probing tasks are given to cover the shallow information of sentences, grammar and semantics. The method is shown validated on six text classification datasets, and classification error rate is improved.
  • References | Supplementary Material | Related Articles

  • Osteoporosis Evaluation Method Based on Multimodal Feature Fusion
  • LUO Tao, LI Jian-feng, HAN Jia-hui, WANG Yi-ning, LEI Lu
  • Journal of Beijing University of Posts and Telecommunications. 2019, 42(6): 84-90. DOI:10.13190/j.jbupt.2019-150
  • Abstract ( 745 )     HTML     　　
  • Aiming at the problems that the problems of single diagnosis and low accuracy in the existing osteoporosis assessment, considering the bone image data and questionnaire data, a multi-modal feature fusion osteoporosis evaluation method based on deep neural network was proposed. And, for the characteristics of shallow image and fixed structure of bone image, Unet is used to perform image segmentation preprocessing to remove redundant information. In view of the shortcomings of ordinary convolution operations in grasping the global information, a new convolutional neural network based on non-local module was proposed to further enrich the feature information. Cross-validation shows that the proposed multimodal feature fusion method has obvious advantages compared with the machine learning method using only image data or questionnaire data alone, and the classification accuracy rate is increased by 3.2% and 22.3%.
  • References | Supplementary Material | Related Articles

  • A Data Driven Characteristically Filtering Method for 3D Flow Field
  • XIONG Guang-zheng, HUANG Zhi-bin, DAI Zhi-tao, YANG Wu-bing
  • Journal of Beijing University of Posts and Telecommunications. 2019, 42(6): 91-97. DOI:10.13190/j.jbupt.2019-152
  • Abstract ( 478 )     HTML     　　
  • With the wide application of fluid mechanics, more and more large-scale fine flow fields have emerged. Overlapping streamlines and dense fields make it hard to use the traditional streamline visualization methods to characterizes the flow fields or process with large-scale fine flow fields. Based on the idea of data-driven, this paper presents an algorithm to implement the characterization of large-scale fine flow fields. The algorithm characterizes streamlines obtained by widely spreading seed points, calculates the features of each point, segments the streamlines based on the features, and then constructs a set of feature vectors and a set of word vectors. Then, the algorithm calculates the geometric feature similarity between streamlines to evaluate streamline similarity and achieves streamlines filtering. Two typical application scenarios, streamline query and flow field compression, verify the proposed method.
  • References | Supplementary Material | Related Articles

  • A Key Variable Detection Algorithm in Multivariate Manufacturing Process
  • YU Kai-xiang, CHEN Zhen-hao, ZHANG Si-hai
  • Journal of Beijing University of Posts and Telecommunications. 2019, 42(6): 98-104. DOI:10.13190/j.jbupt.2019-172
  • Abstract ( 576 )     HTML     　　
  • A key variable detection algorithm based on machine learning in multivariate manufacturing process was proposed. It uses the machine learning classifier to mathematically model the multivariate manufacturing process. And the performance change of the classifier after the process variable shuffled randomly is used as an evaluation index to detect the key variables that lead to relatively abnormal product quality. Simulation data of the multivariate manufacturing process is designed and generated. The performance of the algorithm was verified by simulation data set and two actual production case data sets based on a factory in China. Both verifications show that the algorithm has good detection performance.
  • References | Supplementary Material | Related Articles

  • SA-Siam++: the Two-Branch Siamese Network-Based Object Tracking Algorithm
  • TIAN Lang, HUANG Ping-mu, Lü Tie-jun
  • Journal of Beijing University of Posts and Telecommunications. 2019, 42(6): 105-110. DOI:10.13190/j.jbupt.2019-141
  • Abstract ( 729 )     HTML     　　
  • To deal with the problem of low robustness of SiamFC in complex scenarios in which the object is moving fast, the background is similar to the foreground, and the illumination is strong, a new tracking method called SA-Siam++ was proposed based on two-branch siamese network, including semantic branch which is used to extract semantic information through the hourglass-channel attention mechanism and the appearance branch which is used to extract appearance information through SiamFC. In addition, replacing the AlexNet network with an improved VGG-16 network can significantly increase the feature extraction capabilities. Finally, experiments were carried out on OTB-2013, OTB-2015, UAV123 and VOT2018 which are standard object tracking datasets. It is shown show that the obtained with the proposed algorithm are greatly improved compared with the existing mainstream algorithms, and the average frame rate reaches 49 FPS, that can meet the real-time requirements.
  • References | Supplementary Material | Related Articles

  Reports

  • Spam Web Detection Based on Hybrid-Sampling and Genetic Algorithm
  • LIU Han
  • Journal of Beijing University of Posts and Telecommunications. 2019, 42(6): 111-117. DOI:10.13190/j.jbupt.2019-147
  • Abstract ( 475 )     HTML     　　
  • Spam web detection is of ten troubled by the problem of unbalanced data and high feature space dimension. In order to solve these two problems, the ensemble classification algorithm based on random hybrid-sampling and genetic algorithm was proposed. Firstly, a number of balanced training data subsets is obtained by reducing the number of majority samples through random sampling and generating minority samples by synthetic minority over-sampling technique(SMOTE) method. Then, the improved genetic algorithm is used to reduce the dimension of training data set to obtain multiple subsets of training data with optimal feature. Extreme gradient boosting(XGBoost)is also used as the classifier to train multiple balanced data subsets, and so a new classifier is obtained by ensemble multiple classifiers with simple voting method. Finally, the test set is predicted and the final prediction is obtained. Experiments show that, compared with XGBoost, the proposed algorithm improves the accuracy by about 19.25%, reduces the time to build the learning model, and improves the classification performance.
  • References | Supplementary Material | Related Articles

  • Implicit Authentication Mechanism Based on Context Awareness for Smartphone
  • WANG Ren-zhong, TAO Dan
  • Journal of Beijing University of Posts and Telecommunications. 2019, 42(6): 118-125. DOI:10.13190/j.jbupt.2019-043
  • Abstract ( 661 )     HTML     　　
  • Considering that the existing implicit authentication method for smartphone is difficult to be applied in practice, an implicit authentication mechanism based on context awareness was proposed. Firstly, the tapping behavior data acquired by embedded sensors (i.e. accelerometer, gyroscope, and magnetometer) of the smartphone is used to identify the context information (that is, the user's body posture state). Then, the context feature and tapping feature is extracted from the data after noise elimination which is employed to train the corresponding authentication sub-model for each context. Experiment evaluation is performed using the collected 7000+ data. It is shown that the average false acceptance rate (FAR) and false rejection rate (FRR) of the proposed scheme for different users is 1.29% and 1.03% respectively. Compared with the authentication method without context awareness, FAR and FRR is reduced by 1.72% and 2.59% respectively. The proposed mechanism can effectively improve the accuracy of authentication.
  • References | Supplementary Material | Related Articles

  • Unanswerable Questions Recognition by Semantic Discrepancy Detection
  • LIU Yong-bin, WANG Xiao-jie, YUAN Cai-xia, YI Lian
  • Journal of Beijing University of Posts and Telecommunications. 2019, 42(6): 126-133,141. DOI:10.13190/j.jbupt.2019-202
  • Abstract ( 512 )     HTML     　　
  • Machine reading comprehension (MRC) with unanswerable questions is challenging to the field of natural language processing research. Unlike previous work which ignores the mechanism of answerable and unanswerable, the semantic conflicts detection-based MRC network (SCDNet) was proposed aiming at detections of no-answer (NA) questions through semantic conflicts detection network. The basic idea is that if the given question is unanswerable, there exists semantic absence or conflicts between the question and the reference passages. Therefore, SCDNet predicts the NA probability by checking whether the passage covers the integral semantics of the question. Besides, in order to extract the exact answer from the passage, SCDNet is applied an answer length penalty in the loss function, which helps the learning objective to be more consistent with the evaluation metrics. SCDNet packs the NA question predictor and the answer extractor in a joint model and is trained in an end-to-end manner. Experiments show that SCDNet performs better than some strong baseline models, and achieve an F1 score of 72.43 and 76.96 NA accuracy on SQuAD 2.0 dataset.
  • References | Supplementary Material | Related Articles

  • Label Relevance Based Multi-Label Scratch Classification Algorithm
  • PENG Cong, SUN Yan, QI Peng
  • Journal of Beijing University of Posts and Telecommunications. 2019, 42(6): 134-141. DOI:10.13190/j.jbupt.2019-126
  • Abstract ( 512 )     HTML     　　
  • In order to implement the classification of projects in visual programming field of Scratch, a multi-label classification algorithm (MLLR) appears based on label relevance. An effective multi-label classification model for Scratch projects was constructed. Firstly, the block usage features, the computational thinking skill features and the Halstead features of projects are extracted as classification features. Then, the RAKEL algorithm randomly chooses label subsets, ignoring the relevance between labels, thereafter an improved MLLR algorithm was proposed. This method divides label subsets according to the relevance between multiple labels, and then trains the corresponding label power set sub-classifiers. Experiments show that MLLR algorithm is superior to RAKEL and other multi-label classification algorithms in classification performance and time performance, The classification model constructed has a strong applicability for Scratch projects, and the accuracy of classification reaches 81.3%.
  • References | Supplementary Material | Related Articles

  • Personalized Hierarchical Recurrent Model for Session-Based Recommendation Systems
  • WANG Ya-qing, GUO Cai-li, CHU Yun-fei, ZHOU Hong-hong, FENG Chun-yan
  • Journal of Beijing University of Posts and Telecommunications. 2019, 42(6): 142-148. DOI:10.13190/j.jbupt.2019-143
  • Abstract ( 498 )     HTML     　　
  • The existing studies of session-based recommendations mainly focus on the short-term and long-term interests of users. In order to accurately depict behavior patterns of users, the author introduces the medium-term interests and proposes personalized hierarchical recurrent model (PHRM) based on recurrent neural networks (RNNs), to learn a comprehensive description of user interests by jointly leveraging session, block and global behaviors in a unified framework. First, to model short-term interests, a session-level RNN is designed to capture sequential patterns in sessions. Next, to further describe medium-term interests, a block-level RNN is added to capture correlations across sessions in a block. Then, a user-level RNN is devised to track evolution of long-term interests. Finally, the article designs fusion layers with different interaction mechanisms to effectively integrate cross-level interest information. Simulations on three real-world datasets show that PHRM outperforms the state-of-the-art recommendation methods, with Recall@10 increasing by 18.35%.
  • References | Supplementary Material | Related Articles

  • Location Prediction Model Based on User Behavior Sequence Features
  • HU Zheng, LIU Yi-shan, ZHU Xin-ning, YU Jian-gang
  • Journal of Beijing University of Posts and Telecommunications. 2019, 42(6): 149-154. DOI:10.13190/j.jbupt.2019-106
  • Abstract ( 599 )     HTML     　　
  • In order to solve the problem of ignoring the character of user behavior sequence and limiting the improvement of prediction accuracy, two location prediction models based on the character of user behavior sequence were proposed. Firstly, behavior+context+profile+RNN (BCP-RNN) model is constructed by manually extracting sequence features of user behaviors and integrating the features into the location prediction model. Then three-layer symmetrical neural network (TS-RNN) model is constructed by automatically learning behavior sequence features based on the recurrent structure of RNN model and integrating the features into location prediction model. Experiments show that, compared with the existing location prediction models, BCP-RNN and TS-RNN improves the prediction performance, verifying the importance of behavior sequence features in mining user movement patterns. Besides, compared with the BCP-RNN model of manually extracting behavior sequence features, TS-RNN not only saves the cost of artificial feature extraction, but also makes up for the deviation caused by one-sided human analysis, and has higher prediction accuracy.
  • References | Supplementary Material | Related Articles

  • Paragraph Image Captioning with Deep Fully Convolutional Neural Networks
  • LI Rui-fan, LIANG Hao-yu, FENG Fang-xiang, ZHANG Guang-wei, WANG Xiao-jie
  • Journal of Beijing University of Posts and Telecommunications. 2019, 42(6): 155-161. DOI:10.13190/j.jbupt.2019-057
  • Abstract ( 516 )     HTML     　　
  • How to improve the coherence among descriptive sentences for the paragraph image captioning is paid attention currently. A fully convolutional neural architecture for paragraph image captioning was proposed. An image representation is first obtained using a region detector based on a convolutional network. Then a hierarchical deep convolutional decoder is constructed to translate the image representation, automatically generating a paragraph text description. In addition, the gating mechanism is embedded in the convolutional decoder network to improve memory capacity of the model. Experiments demonstrate that compared with those traditional methods based on recurrent neural networks, the proposed algorithm can generate more coherent paragraph text descriptions for images, achieving better results on evaluation metrics.
  • References | Supplementary Material | Related Articles

  • Prediction of PM_2.5 Concentration Based on Ensemble Learning
  • PENG Yan, ZHAO Zi-ru, WU Ting-xian, WANG Jie
  • Journal of Beijing University of Posts and Telecommunications. 2019, 42(6): 162-169. DOI:10.13190/j.jbupt.2019-153
  • Abstract ( 897 )     HTML     　　
  • The increase of PM2.5 is a cause of haze. Effectively predicting PM2.5 concentration and analyzing its influence factors play an important role in air quality forecasting and controlling. Considering nonlinearity and uncertainty of PM2.5 concentration, a PM2.5 concentration prediction model which firstly selects features using integrated trees was presented based on ensemble trees-gradient boosting decision tree(GBDT). With standard arithmetic mean aggregation method, the article calculates the influence degree of each feature on the increment of PM2.5 concentration, and provides the impact ranking from strong to weak. The grid-search to select the optimal parameters of the GBDT algorithm was used, such as the depth of the tree. Two datasets, the pollutant concentration data and meteorological observation data of Beijing from 2015 to 2016, are used in the prediction model proposed. Compared with standard models such as decision tree, random forest and support vector machine, the ensemble trees-GBDT model is found to be lower mean absolute errors, lower root mean square errors and better generalization ability.
  • References | Supplementary Material | Related Articles

  • Sketch Based Face Recognition in Video
  • CHEN Ping, HUANGFU Da-peng, WANG Xing-jian, DANG De-peng
  • Journal of Beijing University of Posts and Telecommunications. 2019, 42(6): 170-176. DOI:10.13190/j.jbupt.2019-183
  • Abstract ( 496 )     HTML     　　
  • For public securities, the sketch method is often need to be retrieved in video for seeking key people. An improved recognition algorithm that only extracts global or local features from sketch and video media data was presented. The problem of face recognition is turned into face retrieval. The above two medias data is represented as the so called third kind of data, i.e. image list and the final feature comparison is completed by comparing image list. The system designed here extends the research scope of face recognition and supports the retrieval of multi-media data. Compared with face recognition of face video analysis combined system for sketch-video data, the proposed design improves the correct recognition rate of precision and area under the curve, while the equal error rate of equal error rate decreases greatly.
  • References | Supplementary Material | Related Articles