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....
 
 
2019 Vol.32 Issue.10, Published 2019-10-25
Papers and Reports    Researches and Applications   
   
Papers and Reports
869 Scale-Aware Partition-Based Cooperative Correlation Filter Tracking Algorithm
CHEN Can, CHEH Zhaojiong, GU Yang, YE Dongyi
When partition-based correlation filtering tracking algorithm is applied to deal with scale changes and target occlusion problems, the estimation of local sub-block state tracking and the relationship between local sub-block and scale change are inaccurate.To address this issue, a scale-aware partition-based cooperative correlation filter tracking algorithm is proposed. A method of local sub-block occlusion discrimination based on time-sequence smooth constraint is adopted, and the scoring strategy of the existing algorithm is improved. A cooperative motion strategy for sub-blocks is then designed to make the occluded or the deformed sub-blocks follow the normal ones to their due positions. And the ratio between target scale and distributed location of sub-blocks aggregation and dispersion is discovered to perceive target scale changes and estimate sizes. Experiments indicate that the proposed algorithm achieves better performance.
2019 Vol. 32 (10): 869-881 [Abstract] ( 436 ) [HTML 1KB] [ PDF 2569KB] ( 291 )
882 Hyperspectral Image Classification Based on 3D Multi-scale Feature Fusion Residual Network
GUO Wenhui, CAO Feilong
Hyperspectral image(HSI) data used for training in deep learning are insufficient, and therefore deeper network is unfavorable for spectral-spatial feature extraction. To solve this problem, a 3D multi-scale feature fusion residual network is proposed. Spectral-spatial features are learned by deep learning and multi-scale feature fusion. Firstly, the dimension of 3D-HSI data is adaptively reduced, and the images after dimensionality reduction are used as the input of the network. Secondly, spectral-spatial features are extracted successively through multi-scale feature fusion residual blocks and features of different scales are fused. The information flow is enhanced through sharing features and richer features are obtained. Finally, the network is trained end-to-end and tested on corresponding datasets. Experimental results show the satisfactory classification performance of the proposed network.
2019 Vol. 32 (10): 882-891 [Abstract] ( 492 ) [HTML 1KB] [ PDF 2094KB] ( 298 )
892 Temporal Group Deep Network Action Recognition Algorithm Based on Attention Mechanism
HU Zhengping, DIAO Pengcheng, ZHANG Ruixue, LI Shufang, ZHAO Mengyao
Inspired by the mechanism of human visual perception, a temporal group deep network action recognition algorithm based on attention mechanism is proposed under the framework of deep learning. Aiming at the deficiency of local temporal information in describing complex actions with a long duration, the video packet sparse sampling strategy is employed to conduct video level time modeling at a lower cost. In the recognition stage, channel attention mapping is introduced to further utilize global feature information and capture classified interest points, and channel feature recalibration is performed to improve the expression ability of the network. Experimental results on UCF101 and HMDB51 datasets show that the recognition accuracy of the proposed algorithm is high.
2019 Vol. 32 (10): 892-900 [Abstract] ( 440 ) [HTML 1KB] [ PDF 1180KB] ( 428 )
901 Human Action Recognition Using RGB-D Image Features
TANG Chao, WANG Wenjian, ZHANG Chen, PENG Hua, LI Wei
Since the existing multi-modal feature fusion methods cannot measure the contribution of different features effectively, a human action recognition method based on RGB-depth image features is proposed. Firstly, the histogram of oriented gradient feature based on RGB modal information, the space-time interest points feature based on depth modal information, and the joints relative position feature based on joints modal information are acquired to express human actions, respectively. Then, nearest neighbor classifiers with different distance measurement formulas are utilized to classify prediction samples expressed by the three modal features. The experimental results on public datasets show that the proposed method is simple, fast and efficient.
2019 Vol. 32 (10): 901-908 [Abstract] ( 653 ) [HTML 1KB] [ PDF 812KB] ( 398 )
909 Semi-supervised Preference Learning Algorithm
ZHAO Min, LIU Jinglei

To improve low robustness and reconstruction accuracy of recommendation system, a semi-supervised preference learning algorithm is proposed to obtain potential preferences via preference learning and implement recommendations. The l2,1 norm is utilized as the regularization of the optimization objective function to eliminate the noises and outliers. The graph Laplacian regularization is employed to integrate the side information of UI matrix to realize multi-image fusion and improve recommendation precision. The experiments on Movielens 10M and Netflix datasets indicate that the proposed algorithm produces high precision, speed and robustness.
2019 Vol. 32 (10): 909-916 [Abstract] ( 352 ) [HTML 1KB] [ PDF 536KB] ( 380 )
Researches and Applications
917 Image Patch Transform Training and Non-convex Regularization for Image Denoising and Deblurring
YANG Ping, ZHAO Yanwei, ZHENG Jianwei, WANG Wanliang
Aiming at insufficient sampling of image patches in the process of over-complete dictionary training of sparse representation model, an algorithm of image patch transform training and non-convex regularization for image denoising and deblurring is proposed. The image patch search strategy with inter-group variance constraint is adopted, and the selected dictionary set is transposed and learned according to the adaptive soft threshold. The lp(0<p<1) norm is adopted in the reconstruction process to ensure strong sparsity of the results. Split Bregman method is employed to solve the proposed non-convex model. Experimental results show that the proposed algorithm produces better visual effect and Denoising and Deblurring effect.
2019 Vol. 32 (10): 917-926 [Abstract] ( 314 ) [HTML 1KB] [ PDF 1211KB] ( 210 )
927 Heterogeneous FPGA Based Convolutional Network Accelerator
ZHOU Xixiong, ZHONG Sheng, ZHANG Weijun, WANG Jianhui
Computational complexity of neural network methods is high, and its application in embedded scenarios is limited. To solve this problem, a convolutional network accelerator based on heterogeneous field programmable gate array is proposed. The sliding window is employed to accelerate the convolution calculation process, and thus the convolution process of different input and output channels can be handled. A 8 bit fixed-point accelerator is designed combining network quantization process, and the usage of computing resources is reduced. Experiments demonstrate that the proposed fixed-point accelerator achieves a higher computing speed and a lower power consumption with a less performance loss.
2019 Vol. 32 (10): 927-935 [Abstract] ( 341 ) [HTML 1KB] [ PDF 792KB] ( 445 )
936 Non-negative Low Rank Graph Embedding Algorithm Based on L21 Norm
LIU Guoqing, LU Guifu, ZHANG Qiang, ZHOU Sheng
In the existing non-negative matrix factorization(NMF) methods, low-dimensional repre-sentation is directly computed on the original high-dimensional image dataset. Besides, NMF methods are sensitive to noise data, noise labels, unreliable graphs and poor in robustness. To solve these problems, a non-negative low rank graph embedding algorithm based on L21 norm(NLGEL21) is proposed. NLGEL21 takes the effective low rank structure and geometric information of the original dataset into account. L21 norm is introduced into the function of graph embedding and data reconstruction to further improve its robustness. In addition, a multiplicative iteration formula and convergence proof for NLGEL21 are produced. Experiments on ORL, CMU PIE and YaleB face databases show the superiority of NLGEL21.
2019 Vol. 32 (10): 936-944 [Abstract] ( 332 ) [HTML 1KB] [ PDF 673KB] ( 386 )
945 Heterogeneous Ensemble Learning Algorithm Based on Label Distribution Learning
WANG Yibin, TIAN Wenquan, CHENG Yusheng
To improve prediction accuracy, a stacking integration framework in machine learning is employed to learn label distribution through multiple classifiers, and a heterogeneous ensemble learning algorithm based on label distribution learning(HELA-LDL) is proposed. A two-layer model framework is constructed, and the sample data are combined through the first layer structure to integrate the learning results of each classifier. Finally, the fusion results are input to the second layer classifier as the original feature, and the labels are predicted to be a label distribution with confidence. Comparative experiments on specialized datasets show that HELA-LDL is superior to other algorithms in various scenes. The stability analysis further illustrates the effectiveness of HELA-LDL.
2019 Vol. 32 (10): 945-954 [Abstract] ( 359 ) [HTML 1KB] [ PDF 941KB] ( 304 )
955 Nonlocal Similarity Based Tensor Train Factorization for Color Image Completion
JIA Huidi, HAN Zhi, CHEN Xiai, TANG Yandong
In data acquisition and transformation, the data are more or less lost. Therefore, the results of computer vision tasks such as object recognition and tracking are affected. In a natural image, there are many similar structures and patterns with repeated features. With these similar structures and patterns, a method of nonlocal similarity based tensor train factorization for color image completion is proposed. Nonlocal similarity of images are employed to exploit the low rank feature, and modeling is conducted by tensor train factorization to further mine low rank information through transforming a low-order tensor to higher-order one. Experimental results validate the proposed method in image completion.
2019 Vol. 32 (10): 955-963 [Abstract] ( 462 ) [HTML 1KB] [ PDF 2521KB] ( 349 )
964
2019 Vol. 32 (10): 964-964 [Abstract] ( 1048 ) [HTML 1KB] [ PDF 148KB] ( 723 )
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