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....

2020 Vol.33 Issue.2, Published 2020-02-25 Papers and Reports    Researches and Applications        Papers and Reports 95 Attribute Reduction Method Based on Improved Binary Glowworm Swarm Optimization Algorithm and Neighborhood Rough Set PENG Peng, NI Zhiwei, ZHU Xuhui, XIA Pingfan Aiming at the problems of dimension reduction and redundancy removing, an attribute reduction method based on improved binary glowworm swarm optimization algorithm and neighborhood rough set is proposed. Firstly, the population is collaborative initialization using reverse learning, and the mapping of the change function based on Sigmoid is employed for binary coding, and an improved binary glowworm opti-mization algorithm is proposed with Lévy flight position update strategy. Secondly, neighborhood rough set is employed as an evaluation criterion, and the proposed algorithm is utilized as an search strategy for attribute reduction. Finally, experiments on the standard UCI datasets demonstrate the effectiveness of the attribute reduction method, and the better convergence speed and accuracy of the proposed algorithm is verified. 2020 Vol. 33 (2): 95-105 [Abstract] ( 560 ) [HTML 1KB] [ PDF 805KB] ( 299 ) 106 Label-Enhanced Reading Comprehension Model SU Lixin, GUO Jiafeng, FAN Yixing, LAN Yanyan, CHENG Xueqi In the existing extractive reading comprehension models, only the boundary of answers is utilized as the supervision signal and the labeling processed by human is ignored. Consequently, learned models are prone to learn the superficial features and the generalization performance is degraded. In this paper, a label-enhanced reading comprehension model is proposed to imitate human activity. The answer-bearing sentence, the content and the boundary of the answer are learned simultaneously. The answer-bearing sentence and the content of the answer can be derived from the boundary of the answer and these three types of labels are regarded as supervision signals. The model is trained by multitask learning. During prediction, the probabilities from three predictions are merged to determine the answer, and thus the generalization performance is improved. Experiments on SQuAD dataset demonstrate the effectiveness of LE-Reader model. 2020 Vol. 33 (2): 106-112 [Abstract] ( 458 ) [HTML 1KB] [ PDF 771KB] ( 229 ) 113 Multi-layers Context Convolutional Neural Network for Object Detection WANG Hao, SHAN Wenjing, FANG Baofu Insufficient feature information in object detection results in low accuracy of small targets and occluded targets detection. Therefore, multi-layers context convolutional neural network (MLC-CNN) is proposed, and contextual information of multiple layers is extracted to combine local features of objects in object detection. MLC-CNN consists of region proposal network (RPN) sub-network and multi-layers context (MLC) sub-network. RPN sub-network is employed to capture feature vectors with the fixed length as object features, and MLC is employed to obtain the corresponding contextual information of the different feature maps. Finally, two kinds of information are fused. In addition, hard example training is employed to solve the problem of imbalance data. Experiments on PASCAL VOC2007 and PASCAL VOC2012 datasets indicate that mean average precision (mAP) value is improved. 2020 Vol. 33 (2): 113-120 [Abstract] ( 657 ) [HTML 1KB] [ PDF 733KB] ( 490 ) 121 Preferred Strategy Based Self-adaptive Ant Lion Optimization Algorithm LIU Jingsen, HUO Yu, LI Yu Ant lion optimization (ALO) algorithm produces low convergence speed and accuracy in high dimensional solution and it is inclined to fall into local extremum. Therefore, a preferred strategy based self-adaptive ant lion optimization algorithm (PSALO) is proposed. The adaptive boundary mechanism is introduced into the process of ant walking around the ant lion to increase the activity of ant population and prevent the algorithm from falling into the local extremum. The optimal roulette strategy is added in the ant lion selection by roulette to maintain the diversity of ant lion individuals and accelerate the convergence speed of the algorithm. The dynamic proportional factor is added into the ant position update formula to improve the exploration ability of the algorithm in the early stage and the development ability in the later stage. Theoretical analysis proves that the time complexity of the proposed algorithm is same as that of ALO. Optimized simulation experiment of 16 standard test functions with different features in multiple dimensions indicates good feasibility of the proposed algorithm. The optimization precision and convergence speed are improved significantly and they are less affected by the dimension variation. The ability in high dimensional solution is better and more stable. 2020 Vol. 33 (2): 121-132 [Abstract] ( 617 ) [HTML 1KB] [ PDF 914KB] ( 317 ) 133 Video Anomaly Detection Algorithm Combining Global and Local Video Representation HU Zhengping, ZHAO Mengyao, XIN Bingyi Aiming at the problem of video anomaly detection, a video anomaly detection algorithm combining global and local video representation is proposed. Firstly, the input video continuous multi-frames are divided into video blocks. The video blocks are divided into non-overlapping space-time cubes according to the spatial position. The global spatiotemporal grid position support vector data description (SVDD) model based on spatial position is constructed using the space-time cubes motion features. Then, the local texture motion features are extracted for the moving targets of videos. SVDD algorithm is utilized to obtain the hypersphere boundary around the target features, and the normal behavior model of the moving targets is constructed. Finally, the two parts are combined to conduct more comprehensive detection. Experiments on public datasets verify the effectiveness of the proposed algorithm. 2020 Vol. 33 (2): 133-140 [Abstract] ( 329 ) [HTML 1KB] [ PDF 909KB] ( 281 ) Researches and Applications 141 Chinese Paintings Sentiment Recognition via CNN Optimization with Human Cognition SHENG Jiachuan, CHEN Yaqi, WANG Jun, LI Liang Most existing methods are lack of analysis of Chinese paintings sentiment. In this paper, an algorithm of Chinese paintings sentiment recognition via convolutional neural network (CNN) optimization with human cognition is proposed. Firstly, the region of interest from painting is extracted to obtain the square area with rich emotional expression according to image saliency and brushstroke complexity. Then, the deep learning network structure is optimized by combining feature visualization and the knowledge of emotional expression techniques in traditional Chinese paintings. Finally, a rebuilt network is fine-tuned for Chinese paintings sentiment classification task. In the experiment, 1000 Chinese paintings are classified with four emotions and the accuracy rate is better than that of other convolutional neural networks. The effect of model operation is explained through ablation and visualization experiments, and thus the ability of the proposed algorithm to recognize the sentiment of traditional Chinese paintings is confirmed. 2020 Vol. 33 (2): 141-149 [Abstract] ( 398 ) [HTML 1KB] [ PDF 1719KB] ( 398 ) 150 Deep Incremental Image Classification Method Based on Double-Branch Iteration HE Li, HAN Keping, ZHU Hongxi, LIU Ying To solve the catastrophic forgetting problem caused by incremental learning, a deep incremental image classification method based on double-branch iteration is proposed. The primary network is utilized to store the acquired old class knowledge, while the branch network is exploited to learn the new class knowledge. The parameters of the branch network are optimized by the weight of the primary network in the incremental iteration process. Density peak clustering method is employed to select typical samples from the iterative dataset and construct retention set. The retention set is added into the incremental iteration training to mitigate catastrophic forgetting. The experiments demonstrate the better performance of the proposed method. 2020 Vol. 33 (2): 150-159 [Abstract] ( 393 ) [HTML 1KB] [ PDF 1187KB] ( 292 ) 160 Joint Hashing Feature and Classifier Learning for Cross-Modal Retrieval LIU Haoxin, WU Xiaojun, YU Jun To solve the problem of low retrieval accuracy and long training time in cross-modal retrieval algorithms, a cross-modal retrieval algorithm joining hashing feature and classifier learning (HFCL) is proposed. Uniform hash codes are utilized to describe different modal data with the same semantics. In the training stage, label information is utilized to study discriminative hash codes. And the kernel logistic regression is adopted to learn the hash function of each modal. In the testing stage, for any sample, the hash feature is generated by learned hash function, and another modal datum related to its semantics is retrieved from the database. Experiments on three public datasets verify the effectiveness of HFCL. 2020 Vol. 33 (2): 160-165 [Abstract] ( 404 ) [HTML 1KB] [ PDF 614KB] ( 381 ) 166 Collaborative Filtering Hybrid Filling Algorithm for Alleviating Data Sparsity REN Yonggong, WANG Siyu, ZHANG Zhipeng The rating matrix is sparse, and the traditional user-based collaborative filtering cannot provide high-precision satisfactory recommendations for target users. Based on this situation, a hybrid filling collaborative filtering (HFCF) is proposed to alleviate the problem of data sparsity. From the perspective of the item, the sparse matrix is filled according to the rating information of the similar items. And from the viewpoint of users, the neighborhood of the target users is calculated according to the filled matrix. The items with the largest number of common ratings are selected to further fill the matrix. Experiments on two real datasets indicate that the proposed algorithm effectively improves the recommendation precision and relieves the data sparsity problem without any additional information. 2020 Vol. 33 (2): 166-175 [Abstract] ( 431 ) [HTML 1KB] [ PDF 606KB] ( 237 ) 176 Age Estimation Algorithm Based on Deep Cost Sensitive CNN LI Daxiang, MA Xuan, REN Yaqiong, LIU Ying Aiming at the problems of the imbalance of sample size in age estimation and the cost of misclassification between different classes, the cost sensitivity is embedded into the deep learning framework, and an age estimation algorithm based on deep cost sensitive convolutional neural networks (CNN) is proposed. Firstly, a loss function is established for each age category to solve the imbalance problem of the training samples. Then, a cost vector is defined to reflect the cost difference caused by misclassification between different classes, and an inverse cross entropy error function is constructed. Finally, the above methods are merged to derive a loss function for CNN to learn the robust face representation for age estimation during the training process. Experiments on different age estimation standard image sets verify the effectiveness of the proposed algorithm. 2020 Vol. 33 (2): 176-181 [Abstract] ( 471 ) [HTML 1KB] [ PDF 695KB] ( 417 ) 182 Stereo Matching Algorithm Based on Multiscale Fusion XU Xuesong, WU Junjie Aiming at the problems of local stereo matching methods, such as difficulty in selecting sizes of matching windows and low accuracy of stereo matching in weak texture or highlight region, a multi-scale fusion stereo matching method is proposed by combining convolutional neural network model (CNN) and image pyramid method in this paper. By training CNN, image features of the matched image pairs are learned automatically to complete the calculation of matching cost. Based on the construction of image pyramids, the matched image pairs are expressed in multiple scale. Grounded on the template construction of weak texture region, the matching images of each layer are divided into weak texture region and rich texture region. The image of weak texture region is transformed into small-scale image to calculate the matching degree and reduce the mismatching rate of weak texture image. Then, the image is transformed back to large-scale images and fused with the matching results of rich texture regions to maintain the matching accuracy. Experiments on KITTI dataset indicate that the proposed algorithm yields a better image matching result. 2020 Vol. 33 (2): 182-187 [Abstract] ( 544 ) [HTML 1KB] [ PDF 1223KB] ( 404 )

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