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....

2023 Vol.36 Issue.4, Published 2023-04-25 Papers and Reports    Researches and Applications    Surveys and Reviews        Papers and Reports 287 Self-Supervised Heterogeneous Graph Neural Network Model Based on Collaborative Contrastive Learning of Topology Information and Attribute Information LI Chao, SUN Guoyi, YAN Yeyu, DUAN Hua, ZENG Qingtian The complex structure and rich semantics of heterogeneous graphs can be fully explored by heterogeneous graph neural network models. However, there is mutual interference between attribute information and topology information in the model construction process, resulting in weakened expression capability. Therefore, a self-supervised heterogeneous graph neural network model based on collaborative contrastive learning of topology information and attribute information is proposed. Firstly, the representation of the target nodes is learned from both topological and attribute perspectives. Then, the collaborative contrastive algorithm is employed to optimize the node representation from both perspectives, reducing the interference between topology information and attribute information. Additionally, a positive sample generation method combining the number of meta-paths and node topology similarity is proposed in the self-supervised training process of the model. The experiments on real datasets demonstrate the superior performance of the proposed model. The model code can be found at https://github.com/sun281210/HGTA. 2023 Vol. 36 (4): 287-299 [Abstract] ( 730 ) [HTML 1KB] [ PDF 917KB] ( 639 ) 300 Weight Adaptive Generative Adversarial Imitation Learning Based on Noise Contrastive Estimation GUAN Weifan, ZHANG Xi The traditional imitation learning requires expert demonstrations of extremely high quality. This restriction not only increases the difficulty of data collection but also limits application scenarios of algorithms. To address this problem, weight adaptive generative adversarial imitation learning based on noise contrastive estimation(GLANCE) is proposed to maintain high performance in scenarios where the quality of expert demonstration is inconsistent. Firstly, a feature extractor is trained by noise contrastive estimation to improve the feature distribution of suboptimal expert demonstrations. Then, weight coefficients are set for the expert demonstrations, and generative adversarial imitation learning is performed on the expert demonstrations after redistribution based on the weight coefficients. Finally, ranking loss is calculated based on the known relative ranking evaluation data and weight coefficients are optimized through gradient descent to improve the data distribution. Experiments on multiple continuous control tasks show that GLANCE only needs to obtain 5% of the expert demonstrations dataset as evaluation data to achieve superior performance while the quality of the expert demonstration is inconsistent. 2023 Vol. 36 (4): 300-312 [Abstract] ( 427 ) [HTML 1KB] [ PDF 1850KB] ( 865 ) 313 3D Point Cloud Semantic Segmentation Network Based on Coding Feature Learning TONG Guofeng, LIU Yongxu, PENG Hao, SHAO Yuyuan Now point cloud semantic segmentation is widely applied in various fields such as autonomous driving and virtual reality. However, the current point cloud semantic segmentation algorithms cannot extract relatively complete spatial structure information, and the information for each point is difficult to explain. To address this deficiency, a 3D point cloud semantic segmentation network based on coding feature learning is proposed. Firstly, the local feature encoder is designed based on the introduction of angle information and the enhanced features to learn more complete local spatial structures and alleviate the problem of misclassification of similar objects. Secondly, mixed pooling polymerization module is designed to aggregate rough features and fine features while ensuring the sorting invariance of point cloud. Finally, the multi-scale feature fusion is adopted to fully utilize the different scale features in the encoding layer and achieve accurate semantic segmentation. The experiment on two large benchmark datasets, S3DIS and SemanticKITTI, demonstrates the superiority of the proposed network. 2023 Vol. 36 (4): 313-326 [Abstract] ( 546 ) [HTML 1KB] [ PDF 2496KB] ( 513 ) Surveys and Reviews 327 A Survey of RGB-T Object Tracking Technologies Based on Deep Learning ZHANG Tianlu, ZHANG Qiang RGB-Thermal(RGB-T) object tracking aims to achieve robust object tracking by utilizing the complementarity of RGB information and thermal infrared data. Currently, there are many cutting-edge achievements in RGB-T object tracking based on deep learning, but there is a lack of systematic and comprehensive review literature. In this paper, the challenges faced by RGB-T object tracking are elaborated, and the current mainstream RGB-T object tracking algorithms based on deep learning are analyzed and summarized. Specifically, the existing RGB-T trackers are divided into object tracking methods based on multi-domain network(MDNet), object tracking methods based on Siamese network and object tracking methods based on discriminative correlation filter(DCF) according to their different baselines. Then, the commonly used datasets and evaluation metrics in RGB-T object tracking tasks are introduced and the existing algorithms are compared on the commonly used datasets. Finally, the possible future development directions are pointed out. 2023 Vol. 36 (4): 327-353 [Abstract] ( 867 ) [HTML 1KB] [ PDF 4635KB] ( 751 ) Researches and Applications 354 Multi-feature Fusion Based Short Session Recommendation Model XIA Hongbin, HUANG Kai, LIU Yuan Most research on session recommendation systems focuses on long session recommendation and neglects short sessions. However, in practice short session information account for majority of the information. Due to the limited information contained in short sessions, it is crucial to learn more diverse user preferences and find similar context sessions accurately from short sessions. Therefore, a multi-feature fusion based short session recommendation model(MFFSSR) is proposed. Firstly, the node features and sequence features of sessions are learned respectively via neighborhood aggregation and recurrent neural networks. Secondly, the custom similarity calculation formula is utilized to retrieve the current user history session and other user sessions as context information, which alleviate the lack of information in short sessions. Next, the location-aware multi-head self-attention network is applied to fully explore the hidden features of sessions. Finally, the model recommends the next item based on the current session of multi-feature fusion. Experiments on two real datasets show that the proposed model is superior in terms of metrics. The code for the proposed model can be found at http://github.com/ScarletHK/MFF-SRR. 2023 Vol. 36 (4): 354-365 [Abstract] ( 365 ) [HTML 1KB] [ PDF 700KB] ( 935 ) 366 Solar Cell Defect Generation Algorithm Combining Multiple Perception Fields and Attention ZHOU Ying, PEI Shenghu, CHEN Haiyong, YAN Yuze Aiming at the problem of insufficient image samples for some certain defects in solar cells, a solar cell defect generation algorithm combining multiple perception fields and attention is proposed. The generated images are utilized to train the defect detection model. Firstly, a generative adversarial network with dual discriminators is constructed, and a global discriminator and a local discriminator focuse on global information and local details, respectively. Secondly, the multiple perception field feature extraction is designed and fused with the improved attention module to form a multiple perception field attention module. The module is utilized in the network structure of both the generator and the discriminator. Finally, structural similarity loss and peak signal-to-noise ratio loss are added to the loss function for generator training, and the generated images are mean filtered. The generation experiments for 3 different scales of defect images on the solar electroluminescence dataset show that the structural similarity and peak signal-to-noise ratio are high. Additionally, after training the YOLOv7 detection model with the generated defect images, the average precision values for all three defects are high. 2023 Vol. 36 (4): 366-379 [Abstract] ( 462 ) [HTML 1KB] [ PDF 2933KB] ( 625 )

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