模式识别与人工智能
Home      About Journal      Editorial Board      Instructions      Ethics Statement      Contact Us                   中文
Pattern Recognition and Artificial Intelligence
22 Judgement and Disposal of Academic Misconduct Article
22 Copyright Transfer Agreement
22 Proof of Confidentiality
22 Requirements for Electronic Version
More....
22 Chinese Association of Automation
22 National ResearchCenter for Intelligent Computing System
22 Institute of Intelligent Machines,Chinese Academy of Sciences
More....
 
 
2022 Vol.35 Issue.1, Published 2022-01-25
Papers and Reports    Researches and Applications    Surveys and Reviews   
   
0
2022 Vol. 35 (1): 0-0 [Abstract] ( 682 ) [HTML 1KB] [ PDF 186KB] ( 335 )
Surveys and Reviews
1 Model-Based Reinforcement Learning in Robotics: A Survey
SUN Shiguang, LAN Xuguang, ZHANG Hanbo, ZHENG Nanning

The model-based reinforcement learning makes robots closer to human-like learning and interaction by learning an environment model and optimizing policy or planning based on the model. In this paper, the definition of robot learning problems is described, and model-based reinforcement learning methods in robot learning are introduced, including mainstream model learning and model utilization methods. The mainstream model learning methods are given including the forward dynamics model, the inverse dynamics model and the implicit model. The model utilization methods are presented including model-based planning, model-based policy learning and implicit planning. The current problems on model-based reinforcement learning are discussed. Aiming at the problems of the robot learning task in reality, the application of model-based reinforcement learning is illustrated and the future research directions are analyzed.
2022 Vol. 35 (1): 1-16 [Abstract] ( 942 ) [HTML 1KB] [ PDF 941KB] ( 946 )
Papers and Reports
17 Anomaly Node Detection Method Based on Variational Graph Auto-Encoders in Attribute Networks
LI Zhong, JIN Xiaolong, WANG Yajie, MENG Lingbin, ZHUANG Chuanzhi, SUN Zhi
Graph neural network provides a method of combining structural information and attribute information for attribute network data mining. However, the current unsupervised attribute network anomaly node detection based on graph auto-encoder cannot capture the sub-features of normal nodes, and the false negative rate is high.Anomaly node detection method based on variational graph auto-encoders is proposed to detect abnormal nodes in attribute networks, containing two encoders and a decoder. A variational auto-encoder model consisting of an encoder and a decoder is designed to reconstruct the original input data. A decoder and the second encoder are utilized to learn the latent representation of the network without abnormal node data. The features of normal nodes are learned by the dual variational auto-encoder, and the reconstruction error is applied as the anomaly measure of nodes. Normal nodes composed of normal node features are identified as normal nodes by taking reconstruction error as the anomaly measurement of nodes. Experiments on real network datasets show that the proposed method is able to detect abnormal nodes in attribute networks effectively.
2022 Vol. 35 (1): 17-25 [Abstract] ( 591 ) [HTML 1KB] [ PDF 925KB] ( 504 )
26 Long Tail Recommendation Method Based on Social Network Information
FENG Chenjiao, SONG Peng, ZHANG Kaihan, LIANG Jiye
In the long-tail recommendation scenario, target users are more likely to trust the recommendation results of the friends with similar interests. Therefore, recommending personalized preferences of friends to target users is conducive to improving the performance of long-tail recommendation methods. Accordingly, how to effectively fuse social network information with rating matrix information is an important issue in long-tail recommendation. In this paper, a long-tail recommendation method based on social network information is designed. From the perspective of information fusion, social network and rating matrix information are utilized to share potential feature vectors of users. The information of friend recommendation is taken as an important factor in the proposed recommendation model. User activity level, item unpopularity level, user-item preference level and friend recommendation behavior are taken as inputs, and variational inference is employed to get relevant unknown parameters to realize accurate prediction. Experiments show that the proposed method can recommend long-tail items effectively with high recommendation accuracy.
2022 Vol. 35 (1): 26-36 [Abstract] ( 410 ) [HTML 1KB] [ PDF 806KB] ( 317 )
37 Stereo Matching Algorithm Based on Control Points, RGB Vector Difference and Gradient Census Transform
WANG Sen, WEI Hui, MENG Lingjiang
Aiming at the low matching precision of weak texture region and boundary in traditional stereo matching algorithms, a stereo matching algorithm based on control points, RGB vector difference and gradient census transform is proposed. Firstly, a row matching algorithm based on dynamic time warping(DTW) is exploited to find the optimal matching path. The path is distorted and aligned to select matching control points. Then, RGB vector difference cost is combined with gradient Census transform cost as the matching cost of non-control points. The robustness of pixels is enhanced by gradient Census transform, and the three-dimensional color information of the image is retained by the RGB vector. Thus the higher matching precision is achieved. The combination of the matching cost of control points and non-control points is regarded as the initial matching cost. The RGB vector difference is utilized to acquire the adaptive window of different texture areas. The cost aggregation is calculated by the algorithm in a horizontal and vertical directions of an adaptive window. The cost aggregation is employed to optimize the initial cost. Multi-step optimization is employed to reduce the disparity error rate.Finally, the proposed algorithm is validated on Middlebury dataset for different regions and the disparity is calculated in the real scene of robot. The calculated disparity is verified for three-dimensional reconstruction based on the principle of 3D imaging. Both theoretical analysis and experimental results show that the proposed algorithm reduces the matching error rate of weak texture region and boundary significantly.
2022 Vol. 35 (1): 37-50 [Abstract] ( 429 ) [HTML 1KB] [ PDF 5574KB] ( 255 )
Researches and Applications
51 Feature Descriptor Enhancement for Loop Detection Based on Metric Learning
HAN Bin, LUO Lun, LIU Xiongwei, SHEN Huiliang
Loop detection is an important part of simultaneous localization and mapping (SLAM). In most of the loop detection algorithms, feature descriptors are extracted from data frames, and loops are searched through the Euclidean distance between the descriptors. However, feature enhancement is not conducted on the extracted feature descriptors. In this paper, an algorithm of feature descriptor enhancement for loop detection based on metric learning is proposed. A lightweight algorithm module is designed to transform the feature space of the descriptors to enhance the distinguishing ability of the descriptors and improve the loop detection performance effectively. Pose and descriptors are combined to establish a triple dataset and thus the problem of fuzzy labels is solved. An idea of expanding the dataset is proposed to solve the problem of significantly insufficient loop samples. Based on triplet loss, the proposed loss function is adapted to the loop detection scene, and it is utilized to train a neural network module for feature space transformation. Experiments on KITTI and NCLT datasets show that the generalization ability of the proposed algorithm is strong.
2022 Vol. 35 (1): 51-61 [Abstract] ( 377 ) [HTML 1KB] [ PDF 1740KB] ( 318 )
62 Network Pruning via Automatic Mending Strategy
SU Qihang, QIAN Yeqiang, YUAN Wei, YANG Ming, WANG Chunxiang
To alleviate the problem of the application of deep neural network being restricted owing to the massive computation resources, many network compression strategies including network pruning are put forward. Most of the network pruning methods based on greedy algorithm include training, pruning and fine-tuning, and therefore the optimal pruned structure cannot be obtained. In this paper, combining the rule-based method and the automatic search method, a network pruning method via automatic mending strategy is proposed. The whole pruning process is comprised of four stages: training, pre-pruning, mending and fine-tuning. The structure of the pre-pruned model is improved in the additional mending stage. Particularly, the neural architecture search is utilized to implement network mending. The search space and an efficient search strategy are designed. The estimation process is accelerated based on the filter ranking of the pre-pruning stage. Experiments show that the proposed method can guarantee the network accuracy in the case of high pruning rate.
2022 Vol. 35 (1): 62-70 [Abstract] ( 332 ) [HTML 1KB] [ PDF 947KB] ( 311 )
71 Image Semantic Segmentation Network Based on Semantic Propagation and Fore-Background Aware
LIU Zhanghui, ZHAN Xiaolu, CHEN Yuzhong
Although image segmentation is widely applied in many fields owing to the assistance of better analysis and understanding of images, the models based on fully convolutional neural networks still engender the problems of resolution reconstruction and contextual information usage in semantic segmentation. Aiming at the problems, a semantic propagation and fore-background aware network for image semantic segmentation is proposed. A joint semantic propagation up-sampling module(JSPU) is proposed to obtain semantic weights by extracting the global and local semantic information from high-level features. Then the semantic information is propagated from high-level features to low-level features for alleviating the semantic gap between them. The resolution reconstruction is achieved through a hierarchical up-sampling structure. In addition, a pyramid fore-background aware module is proposed to extract foreground and background features of different scales through two parallel branches. Multi-scale fore-background aware features are captured by establishing the dependency relationships between the foreground and background features, thereby the contextual representation of foreground features is enhanced. Experiments on semantic segmentation benchmark datasets show that SPAFBA is superior in performance.
2022 Vol. 35 (1): 71-81 [Abstract] ( 527 ) [HTML 1KB] [ PDF 3179KB] ( 557 )
82 Object Tracking Algorithm Based on Accelerated Adaptive Spatial-Temporal Background Aware Correlation Filters
LI Yangxiao, WEI Fuyuan, ZHOU Zhenghua, ZHAO Jianwei
Designing a robust tracking algorithm based on correlation filters is an important research direction in the target tracking. The information of background, space and time is significant for improving the tracking performance of the algorithm. Grounded on the background-aware tracking algorithm, an object tracking algorithm based on accelerated adaptive spatial-temporal background aware correlation filter is proposed by fusing the spatial information, temporal information and the adaptability of the spatial weight matrix. Then,the appearance optimization model is solved by the accelerated alternating direction method of multipliers to obtain the spatial weight matrix and the correlation filter to realize the adaptive tracking. The proposed tracking algorithm enhances the discrimination of the tracker for the object from the background with the background information, spatial information and the adaptive spatial weight. The problem of tracking drifting for the case of target occlusion is alleviated by the temporal-regularization term, and the solving process is speeded up by the accelerated alternating direction method of multipliers. Experiments illustrate that the proposed algorithm produces better tracking results in the cases of target occlusion and background interference.
2022 Vol. 35 (1): 82-91 [Abstract] ( 432 ) [HTML 1KB] [ PDF 5381KB] ( 319 )
模式识别与人工智能
 

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