Home      About Journal      Editorial Board      Instructions      Ethics Statement      Contact Us                   中文

Judgement and Disposal of Academic Misconduct Article Copyright Transfer Agreement Proof of Confidentiality Requirements for Electronic Version More.... Chinese Association of Automation National ResearchCenter for Intelligent Computing System Institute of Intelligent Machines,Chinese Academy of Sciences More....

2022 Vol.35 Issue.3, Published 2022-03-25 Papers and Reports    Researches and Applications        Papers and Reports 195 Cross-Media Fine-Grained Representation Learning Based on Multi-modal Graph and Adversarial Hash Attention Network LIANG Meiyu, WANG Xiaoxiao, DU Junping There are problems of feature heterogeneity and semantic gap between data of different media types in cross-media data search, and social network data often exhibits semantic sparsity and diversity. Aiming at these problems, a cross-media fine-grained representation learning model based on multi-modal graph and adversarial Hash attention network(CMFAH) is proposed to obtain a unified cross-media semantic representation and applied to social network cross-media search. Firstly, an image-word association graph is constructed, and direct and implicit semantic associations between image and text words are mined based on the graph random walk strategy to expand the semantic relationship. A cross-media fine-grained feature learning network based on cross-media attention is constructed, and the fine-grained semantic association between images and texts is learned collaboratively through the cross-media attention mechanism. A cross-media adversarial hash network is constructed, and an efficient and compact cross-media unified hash semantic representation is obtained by the joint cross-media fine-grained semantic association learning and adversarial hash learning. Experimental results show that CMFAH achieves better cross-media search performance on two benchmark cross-media datasets. 2022 Vol. 35 (3): 195-206 [Abstract] ( 736 ) [HTML 1KB] [ PDF 732KB] ( 540 ) 207 Face Sketch-Photo Synthesis Method Based on Multi-residual Dynamic Fusion Generative Adversarial Networks SUN Rui, SUN Qijing, SHAN Xiaoquan, ZHANG Xudong Aiming at the low definition and blurry details in the current face sketch-photo synthesis methods, a face sketch-photo synthesis method based on multi-residual dynamic fusion generative adversarial network is proposed. Firstly, a multi-residual dynamic fusion network is designed. Features are extracted from different dense residual modules and then the residual learning is conducted. Then, the corresponding offsets are generated on the basis of the diverse residual features at different levels. The sampling coordinates of the convolution kernels in different locations are changed according to the offsets. Consequently, the network is focused on important feature adaptively, and geometric detail information and high-level semantic information are integrated effectively without gradual information dropping and redundant information interference. Moreover, a multi-scale perceptual loss is introduced to conduct perceptual comparison on the synthetic images of different resolutions for the regularization of synthetic images from coarse to fine. Experiments on Chinese University of Hong Kong face sketch dataset show that the proposed method produces high-definition images with full detail and consistent color and the synthesized image is closer to real face images. 2022 Vol. 35 (3): 207-222 [Abstract] ( 679 ) [HTML 1KB] [ PDF 10709KB] ( 455 ) 223 Item State Transition Functions and Polytomous Knowledge Structures Based on Procedural Knowledge Learning SUN Xiaoyan, LI Jinjin In the assessment of procedural knowledge, skills refer to the operation paths relevant to the solution of an item. Based on the learning assessment of procedural knowledge, a method of delineating polytomous knowledge structure from the state structure of the item itself is proposed to establish a polytomous assessment system for problem solving. Firstly, the response values are set according to the solution or operation process of each item to obtain the item-specific response value set. The item state space is defined by item state transition function, and the problem space is extended to polytomous case. Then,the conjunctive skill maps are derived from the operation paths, and the polytomous knowledge structures delineated by the conjunctive skill maps are discussed. The results show that the polytomous knowledge structure delineated by a skill map based on the conjunctive model satisfies the item-wise intersection closure. Finally, the algorithm steps of delineating polytomous knowledge structure are given, and the effectiveness of the proposed algorithm is illustrated by an example. 2022 Vol. 35 (3): 223-242 [Abstract] ( 384 ) [HTML 1KB] [ PDF 818KB] ( 495 ) 243 Facial Expression Recognition Combining Self-Attention Feature Filtering Classifier and Two-Branch GAN CHENG Yan, CAI Zhuang, WU Gang, LUO Pin, ZOU Haifeng The expression features extracted by the existing facial expression recognition methods are usually mixed with other facial attributes, which is not conducive to facial expression recognition. A facial expression recognition model combining self-attention feature filter classifier and two-branch generative adversarial network is proposed. Two-branch generative adversarial network is introduced to learn discriminative expression representation, and a self-attention feature filtering classifier is proposed as the expression classification module. The cascaded LayerNorm and ReLU are employed to zero the low activation unit and retain the high activation unit to generate multi-level features. The self-attention is utilized to fuse and output the prediction results of multi-level features, and consequently the influence of noise on the recognition results is eliminated to a certain extent. A sliding module based dual image consistency loss supervised model is proposed to learn discriminative expression representations. The reconstruction loss is calculated by a sliding window and more attention is paid to the details. Finally, experiments on CK+, RAF-DB, TFEID and BAUM-2i datasets show the proposed model achieves better recognition results. 2022 Vol. 35 (3): 243-253 [Abstract] ( 551 ) [HTML 1KB] [ PDF 1153KB] ( 453 ) Researches and Applications 254 Gradient-Based Adversarial Ranking Attack WU Chen, ZHANG Ruqing, GUO Jiafeng, FAN Yixing Ranking competition is prevalent in Web retrieval, and undesirable effects are caused by this adversarial attack behavior. Thus, the study on attack methods is conducive to designing a more robust ranking model.The existing attack methods are recognized by people easily and cannot attack neural ranking models effectively.In this paper, a gradient-based adversarial attack method(GARA) is proposed, including gradient-based word importance ranking, gradient-based adversarial ranking attack and embedding-based word replacement. Given a target ranking model, the backpropagation is firstly conducted based on the constructed ranking-based adversarial attack objective. Then the most important words of a specific document is recognized based on the gradient information. These important words are perturbed in the word embedding space based on the projected gradient descent. Finally, by adopting the counter-fitting technology, the document perturbation is completed by substituting the important word with its synonym which is semantically similar to the original word and nearest to the perturbed word vector.Experiments on MQ2007 and MS MARCO datasets demonstrate the effectiveness of the proposed method. 2022 Vol. 35 (3): 254-261 [Abstract] ( 463 ) [HTML 1KB] [ PDF 706KB] ( 334 ) 262 Word-Pair Relation Learning Method for Aspect Sentiment Triplet Extraction XIA Hongbin, LI Qiang, XIAO Yifei Aspect sentiment triplet extraction is designed to identify aspect items with their sentiment tendencies in a comment and to extract the related opinion items. In most of the existing methods, this type of task is divided into several sub-tasks, and then the task is completed by the pipeline composed of the sub-tasks. However, the methods based on pipeline are affected by error propagation and inconvenience for use in practice. Therefore, a word-pair relation learning method for aspect sentiment triplet extraction is proposed, which transforms the aspect sentiment triplet extraction task into an end-to-end word-pair relation learning task. The method contains a word-pair relation tagging scheme, which can unify word-pair relations in sentences to represent all triplets, and a specially built word-pair relation network to output word-pair relation. Firstly, the sentence is encoded by bidirectional grated recurrent unit and mixed attention. Then, sentence coding is converted into tag probabilities through the attention map transform module. Finally, the triplets are extracted from the result of the word-pair relation tag. In addition, the pre-trained bidirectional encoder representation from transformer is applied to the proposed method. Experiments on four standard datasets show that the proposed method is superior. 2022 Vol. 35 (3): 262-270 [Abstract] ( 488 ) [HTML 1KB] [ PDF 639KB] ( 444 ) 271 A Convolutional Vector Network for 3D Mesh Object Recognition QIU Qilu, ZHAO Jieyu, CHEN Yu The 3D object recognition is usually based on convolutional neural networks. The spatial information of 3D objects is lost due to too many pooling layers used in the process of feature aggregation. To solve the above problem, a convolutional vector network for 3D mesh object recognition is proposed in this paper. Firstly, the surface polynomials are introduced to fit the local mesh of the object, and then the surface shape convolution kernels are clustered. The feature vector of structure-aware is generated by self-measuring the similarity between the local mesh in the object and convolution kernels. The multi-headed attention mechanism module is then employed to achieve feature aggregation from local regions to a larger scale to obtain component-level feature vectors. Finally, the 3D objects are classified through the 3D vector network. The proposed network achieves a high classification accuracy on SHREC10, SHREC11 and SHREC15 datasets. In addition, generalization and robustness of the proposed network are demonstrated through the multi-resolution object comparison experiment and the multi-sampling point comparison experiment. 2022 Vol. 35 (3): 271-282 [Abstract] ( 400 ) [HTML 1KB] [ PDF 962KB] ( 414 ) 283 Dual Supervised Network Embedding Based Community Detection Algorithm ZHENG Wenping, WANG Yingnan, YANG Gui A network embedding based community detection algorithm is easy to fall into local extremes during the independent node embedding or clustering process. Aiming at this problem, a dual supervised network embedding based community detection algorithm(DSNE) is proposed. Firstly, a graph auto-encoder is utilized to gain the embedding of nodes to maintain the first-order similarity of the network. Then, the modularity is optimized to find the communities with nodes tightly connected. The communities with similar nodes in the embedding space are discovered by self-supervised clustering optimization. A mutual supervision mechanism is introduced into DSNE to keep the consistency between the discovered communities in modularity optimization and self-supervised clustering and prevent the algorithm from falling into local extremes. Results of comparative experiments show DSNE exhibits better performance on 4 real complex networks. 2022 Vol. 35 (3): 283-290 [Abstract] ( 541 ) [HTML 1KB] [ PDF 1001KB] ( 314 )

Supervised by China Association for Science and Technology Sponsored by Chinese Association of Automation NationalResearchCenter for Intelligent Computing System Institute of Intelligent Machines, Chinese Academy of Sciences Published by Science Press

Copyright © 2010 Editorial Office of Pattern Recognition and Artificial Intelligence Address: No.350 Shushanhu Road, Hefei, Anhui Province, P.R. China Tel: 0551-65591176 Fax:0551-65591176 Email: bjb@iim.ac.cn Supported by Beijing Magtech 　Email:support@magtech.com.cn