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....

2021 Vol.34 Issue.5, Published 2021-05-25 Special Research on Detection, Discrimination and Tracking of Visual Object        Special Research on Detection, Discrimination and Tracking of Visual Object 385 Adaptive Deep Multi-object Tracking Algorithm Fusing Crowd Density LIU Jinwen, REN Weihong, TIAN Jiandong Multi-object tracking technology cannot well solve the problem of multi-object tracking in the scenarios with objects severely occluded, and therefore an adaptive deep multi-object tracking algorithm fusing crowd density is proposed. Firstly, the crowd density maps and object detection results are fused, and the location and the count information of crowd density maps are utilized to correct the detector results to eliminate missing and false detections. Then, adaptive triplet loss is employed to improve the loss function of the re-identification model and thus the discrimination of the algorithm for the re-identification feature is enhanced. Finally, final tracking results are obtained using the appearance and motion information for objects association. It is verified through the experiments that the proposed algorithm effectively solves the problem of multi-object tracking in severely occluded scenes. 2021 Vol. 34 (5): 385-397 [Abstract] ( 530 ) [HTML 1KB] [ PDF 8530KB] ( 474 ) 398 Medical Image Segmentation via Triplet Interactive Attention Network GAO Chengling, YE Hailiang, CAO Feilong Deep learning produces advantages in solving class imbalance due to its powerful ability to extract features. However, its segmentation accuracy and efficiency can still be improved. A medical image segmentation algorithm via triplet interactive attention network is proposed in this paper. A triplet interactive attention module is designed and embedded into the feature extraction process. The module is focused on features in the channel and spatial dimensions jointly, capturing cross-dimensional interactive information. Thus, important features are in focus and target locations are highlighted. Moreover, pixel position-aware loss is employed to further mitigate the impact of class imbalance. Experiments on medical image datasets show that the proposed method yields better performance. 2021 Vol. 34 (5): 398-406 [Abstract] ( 380 ) [HTML 1KB] [ PDF 1250KB] ( 233 ) 407 Few-Shot Metric Transfer Learning Network for Surface Defect Detection HUANG Jian, ZHENG Chunhou, ZHANG Jun, WANG Bing, CHEN Peng Metric learning method of few-shot learning is introduced into the field of defect detection, and a few-shot learning method based on transfer metric learning is proposed to meet the requirement of deep learning method for a large number of learning samples. In the first stage, the deep network is pre-trained on the large datasets which are open or easy to be obtained. In the second stage, the relevant knowledge learned by the network is transferred to the field of surface defect detection through the metric learning module.Experiments show the feasibility of few-shot learning in defect detection. 2021 Vol. 34 (5): 407-414 [Abstract] ( 919 ) [HTML 1KB] [ PDF 1016KB] ( 471 ) 415 CNN-Based Lightweight Flame Detection Method in Complex Scenes LI Xinjian, ZHANG Dasheng, SUN Lilei, XU Yong The existing fire detection methods rely on high-performance machines, and therefore the speeds on the embedded terminals and the mobile ones are not satisfactory. For most of the detection methods, the speed is low and the false detection rate is high, especially for small-scale fires missed detection problems. To solve these problems, a fire detection method based on you only look once is proposed. Depthwise separable convolution is employed to improve its network structure. Multiple data augmentation and bounding box based loss function are utilized to achieve a higher accuracy. The real-time 21ms fire detection on embedded mobile system is realized through parameter tuning with the detection accuracy ensured. Experimental results show that the proposed method improves accuracy and speed on the fire dataset. 2021 Vol. 34 (5): 415-422 [Abstract] ( 744 ) [HTML 1KB] [ PDF 3185KB] ( 596 ) 423 Micro-Expression Recognition Algorithm Based on 3D Convolutional Neural Network and Optical Flow Fields from Neighboring Frames of Apex Frame ZHANG Xuesen, JIA Jingping The existing micro-expression recognition technologies cannot make full use of the spatiotemporal features near the apex frame. Aiming at this problem, a micro-expression recognition algorithm based on 3D convolutional neural network and optical flow fields from the neighboring frames of the apex frame is proposed. Firstly, the optical flow fields between the adjacent frames before and after the apex frame are extracted. The important spatiotemporal information of micro-expressions are retained while the redundant information is removed and the computation load is reduced. Then, a 3D convolutional neural network is employed to extract the enhanced spatiotemporal features from the optical flow fields and thus the classification is completed. Finally, experiments on three spontaneous micro-expression databases show the proposed algorithm produces a better accuracy. 2021 Vol. 34 (5): 423-433 [Abstract] ( 399 ) [HTML 1KB] [ PDF 1471KB] ( 456 ) 434 Parallel Lane Detection Network Based on Image Sequence ZHU Wei, OU Quanlin, HONG Lidong, HE Defeng The existing lane detection neural networks mainly adopt independent single frame image for detection, and therefore they cannot handle the complex and practical application scenarios, such as short-term occlusion of lane and light and shade changes of ground. To solve these problems, a parallel lane detection network based on image sequence is proposed according to the scene characteristics that the continuous images can be obtained in the normal driving process of vehicles. Firstly, a parallel feature extraction structure is designed. A single frame network with high accuracy is employed to extract the features of the current frame image. A lightweight multi-frame network is designed to extract the features of low resolution multi-frame sequential images. The cyclic neural network module is utilized to fuse the extracted sequential features to obtain multi-frame features. Then, the fusion module of single frame feature and multi-frame feature is designed, and the feature map of the lane line is output through upsampling network. The experimental results show that the objective detection accuracy and subjective effect of the proposed network are significantly improved. 2021 Vol. 34 (5): 434-445 [Abstract] ( 323 ) [HTML 1KB] [ PDF 2100KB] ( 329 ) 446 Group Sparse Representation Based on Feature Selection and Dictionary Optimization for Expression Recognition XIE Huihua, LI Ming, WANG Yan, CHEN Hao To solve the over-fitting problem of recognition model on small sample facial expression database, a group sparse representation classification method based on feature selection and dictionary optimization is put forward. Firstly, the feature selection criterion is proposed, and the complementary features of same class-level sparse mode and different intra-class sparse mode are selected to build a dictionary. Then, the dictionary is learned by maximum scatter difference optimization to reconstruct features without distortion and acquire a high discriminative ability. Finally, the optimized dictionary is combined for group sparse representation classification. Experiments on JAFFE and CK+ databases show that the proposed method is robust to sample reduction with high generalization ability and recognition accuracy. 2021 Vol. 34 (5): 446-454 [Abstract] ( 272 ) [HTML 1KB] [ PDF 730KB] ( 185 ) 455 Cross-View Gait Recognition Method Based on Multi-branch Residual Deep Network HU Shaohui, WANG Xiuhui, LIU Yanqiu Convolution neural network based gait recognition cannot make full use of local fine-grained information. To solve the problem, a cross-view gait recognition method based on multi-branch residual deep network is proposed. The multi-branch network is introduced into convolutional neural network to extract features with different granularity in gait contour sequences. Residual learning and multi-scale feature fusion technology are utilized to enhance the feature learning ability of the network. Experimental results on open-accessed gait datasets CASIA-B and OU-MVLP show that the recognition accuracy of the proposed method is higher than that of the existing algorithms. 2021 Vol. 34 (5): 455-462 [Abstract] ( 281 ) [HTML 1KB] [ PDF 662KB] ( 209 ) 463 Global-Local Feature Extraction Method for Fine-Grained National Clothing Image Retrieval ZHOU Qianqian, LIU Li, LIU Lijun, FU Xiaodong, HUANG Qingsong The low accuracy of fine-grained retrieval of national clothing images is caused by different clothing styles, accessories and patterns of national clothing. To address is problem, a global-local feature extraction method for fine-grained clothing image retrieval is proposed. Firstly, the input image is detected to obtain the foreground, styles, accessories and patterns images based on semantic labels of national clothing. Secondly, a multi-branch feature extraction model based on fully convolutional network is constructed to extract features from clothing images in different regions and obtain convolutional features of global, styles, accessories and patterns. Finally, the preliminary retrieval results are obtained by applying a similarity measure to the global features. Then,re-ranking of the result is performed by the local features of top 50 retrieval results and the query image. The final retrieval results are output by the result of re-ranking. The experimental results on the constructed national clothing image dataset show that the proposed method improves the accuracy of national clothing image retrieval effectively. 2021 Vol. 34 (5): 463-472 [Abstract] ( 372 ) [HTML 1KB] [ PDF 2804KB] ( 240 ) 473 Spatially Abnormal Adaptive Target Tracking JIANG Wentao, LIU Xiaoxuan, TU Chao, JIN Yan The correlation filtering algorithm based on spatial regularization cannot suppress the weight of the background region precisely. The credibility of results is reduced by the factors like occlusion and deformation. A spatially abnormal adaptive target tracking is proposed to solve the problems. Firstly, an adaptive spatial regularization term is introduced, and its weight is initialized by significance detection to realize the spatial adaptability. Secondly, the alternating direction method of multipliers is adopted to reduce the complexity of the algorithm. Finally, a verification score is set in each subsequent frame to calculate the reliability of detection result and analyze the abnormal situation. A dynamic update rate for the target model is set and thus the abnormal adaptability is realized. Experiments on four public datasets show that the proposed algorithm produces a good tracking result in different complex scenes, such as deformation, occlusion and illumination variation, and it basically meets the real-time requirements. 2021 Vol. 34 (5): 473-484 [Abstract] ( 346 ) [HTML 1KB] [ PDF 5545KB] ( 263 )

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