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....
 
 
2016 Vol.29 Issue.1, Published 2016-01-30
Papers and Reports    Researches and Applications   
   
Papers and Reports
1 An Improved Non-local Means Denoising Algorithm
CAI Bin, LIU Wei, ZHENG Zhong, WANG Zengfu
Aiming at the problem of similarity calculation for image block in non-local means (NLM) denoising algorithm, a more accurate block matching algorithm is proposed. In this algorithm, the contribution of rotation is taken into alcount. To obtain the blocks similar to the neighborhood of the given pixel, the related blocks surrounding the given pixel are reordered according to their gray values, and then the pixels in the neighborhood of the given pixel are also reordered in the same way. Finally, the distance between the related blocks and the neighborhood of the given pixel are calculated according to the reordered gray values. The candidate blocks with small distance are selected. Furthermore, the more structurally similar blocks are selected from the candidate blocks. To eliminate the effect of noise, the inputted image is processed by a pre-filtering operation before similarity calculation. Simulation experiments show that compared with the original NLM denoising algorithm, the proposed algorithm has better performances in peak signal-to-noise ratio (PSNR), mean structural similarity (MSSIM) and subjective visual effect. Especially, the proposed algorithm has better denoising performance for the images with lots of noise variance.
2016 Vol. 29 (1): 1-10 [Abstract] ( 1200 ) [HTML 1KB] [ PDF 1362KB] ( 1054 )
11 Fast Scalable Subspace Clustering Algorithm
LIU Bo, XIE Bojun, ZHU Jie, JING Liping, YU Jian
Most existing subspace clustering methods are inefficient for large scale datasets and are unable to handle out-of-sample data. To address these problems, a framework is proposed called two-stage sample selection for subspace clustering (TSSC). TSSC consists of two key components: discriminative collaborative representation (DCR) and multi-scale K nearest neighbors (KNN). DCR is combined with multi-scale KNN for feature mapping, and thus the samples belonging to the same subspace have more consistent representation. To enhance the scalability of the algorithm, multi-scale KNN is reused to select some information points from the new feature space by TSSC. Then, a linear classifier is trained according to the clustering result produced by the pre-selected points. Finally, the rest samples are categorized to obtain the final clustering result. Experiments conducted on the real-world datasets verify the effectiveness of TSSC.
2016 Vol. 29 (1): 11-21 [Abstract] ( 624 ) [HTML 1KB] [ PDF 1268KB] ( 658 )
22 Object Tracking Algorithm Based on Hybrid Particle Filter and Sparse Representation
ZHOU Zhiping, ZHOU Mingzhu, LI Wenhui

To reduce the influence of complex environment like illumination variation, appearance change and partial occlusion during the object tracking in the sequence images, a hybrid particle filter tracking method based on global and local information is proposed. The local binary patterns (LBP) textual feature is introduced into the particle filter algorithm. Through sparse coding target sub-block, the local information is fully used, and the global information is taken into account to determine the position of target in the current frame. During the tracking, the robustness of the tracking algorithm is improved since the template is updated in real time. Experimental results show that the proposed tracking algorithm achieves good results in complex background.
2016 Vol. 29 (1): 22-30 [Abstract] ( 547 ) [HTML 1KB] [ PDF 4785KB] ( 695 )
31 Latent Least Square Regression for Subspace Segmentation
CHEN Xiaoyun, CHEN Huijuan
Subspace segmentation is an efficient tool in high dimensional data clustering. However, the construction of affine matrix and the clustering result are directly affected by missing data and noise data. To solve this problem, latent least square regression for subspace segmentation (LatLSR) is proposed. The data matrix is reconstructed in directions of column and row, respectively. Two re-constructed coefficient matrices are optimized alternately, and thus the information in two directions is fully considered. The experimental results on six gene expression datasets show that the proposed method produces better performance than the existing subspace segmentation methods.
2016 Vol. 29 (1): 31-38 [Abstract] ( 551 ) [HTML 1KB] [ PDF 584KB] ( 941 )
39 Active Learning Algorithm of SVM Combining Tri-training Semi-supervised Learning and Convex-Hull Vector
XU Hailong, LONG Guangzheng, BIE Xiaofeng, WU Tian'ai, GUO Pengsong
The large-scale labeled samples can not be acquired easily and the cost of sample labeling is high. Aiming at these problems, an active learning algorithm of support vector machine (SVM) based on tri-training semi-supervised learning and convex-hull vector is proposed in this paper. Semi-supervised learning and active learning are efficiently combined. Firstly, by calculating the convex-hull vector of the sample set, samples of convex-hull vector which are most likely to be support vectors are selected to be labeled. For the existing active learning, the unlabeled samples are no longer used after the most informative samples are selected to be labeled. Secondly, to salve this problem, semi-supervised learning method-based tri-training is introduced into SVM active learning. Thus, the unlabeled samples with higher confidence level of classifying samples are selected and classified as the training sample set, and the useful information for learning machines in the unlabeled samples is exploited. The experimental results on UCI dataset show that the proposed algorithm achieves higher classification accuracy with less labeled samples and it improves generalization performance and reduces the labeling cost of SVM training.
2016 Vol. 29 (1): 39-46 [Abstract] ( 540 ) [HTML 1KB] [ PDF 556KB] ( 624 )
Researches and Applications
47 Interval-Valued Attributes Based Monotonic Decision Tree Algorithm
CHEN Jiankai, WANG Xin, He Qiang, WANG Xizhao
Some learning algorithms of interval-valued attributes are developed in the disorderly situation. The ordinal relation between condition attributes and decision attributes is not taken into account. In this paper, aiming at the defects of the original algorithms, a monotonic decision tree algorithm is proposed to deal with monotonic classification of interval-valued attributes. The possibility degree is used to determine the order relation of interval-valued attributes, the rank mutual information is utilized to measure the monotonic consistency, and the expanded attributes are selected by maximizing the rank mutual information. Furthermore, unstable cut-points are applied to the construction process of interval-valued attributes decision tree to reduce the computing number of rank mutual information and improve the computational efficiency. The experimental results show that the algorithm improves the efficiency and testing accuracy.
2016 Vol. 29 (1): 47-53 [Abstract] ( 604 ) [HTML 1KB] [ PDF 480KB] ( 673 )
54 Peer Group Stock Market Trend Prediction Algorithm Based on Deep Computing
YAO Hongliang, HONG Jingfan, WANG Hao
Aiming at the deficiency of peer group (PG) algorithm, a peer group generation algorithm based on depth computing is proposed. Firstly, the band similarity between target stock and candidate stock is calculated. Then, grounded on depth calculation of intimacy, correlation and liveness, peer group of the target stock is generated, and it is proved that the quality of peer group generated by depth calculation is superior to that of PG algorithm. Since PG algorithm is unable to predict, autoregressive stock market trend prediction algorithm based on peer group (DPG-AR) is proposed by combining peer group algorithm and autoregression model. Peer group is generated by deep computing. Thus, the weights of peer group members are updated. The target stock trend is prodicted by autoregression model. The effectiveness of DPG-AR is verified in the experiment on Shanghai composite index and the corresponding stock.
2016 Vol. 29 (1): 54-62 [Abstract] ( 468 ) [HTML 1KB] [ PDF 468KB] ( 759 )
63 Face Alignment Algorithm Based on Shape Parameter Regression
PENG Mingchao, BAO Jiao, YE Mao, GOU Qunsen, WANG Mengwei
To improve the efficiency of traditional face alignment algorithms, a face alignment algorithm based on shape parametric regression is proposed. Firstly, face is constrained by face shape space and face shape is depicted by a low dimensional shape parameter. Then, a series of shape parameter regressions is learned under the framework of a two-level shape parameter regression algorithm with the combination of an efficient explicit shape feature index method and multiple random feature selection method. Finally, the alignment face shape is portrayed. By the proposed algorithm, the amount of data storage is reduced and the speed of the face alignment is improved. Experiments on complex dataset show that the proposed algorithm obtains good results. Moreover, it can be applied directly on mobile phones, tablet composters and other low-end devices.
2016 Vol. 29 (1): 63-71 [Abstract] ( 753 ) [HTML 1KB] [ PDF 1227KB] ( 1179 )
72 Curve Registration Method for Maximizing Correlation Coefficient Based on Non-uniform Sampling
ZHANG Wenkai, WANG Wenjian, JIANG Gaoxia
In functional data analysis, two kinds of non-uniform sampling methods for curve registration are put forward to improve the efficiency. Slope based non-uniform sampling (SBNS) method samples according to the slope size of the function curve. Arc length based non-uniform sampling (ALBNS) method samples evenly in the arc length of function curve. Two non-uniform sampling methods sample according to characteristics of curves instead of sampling evenly in the time axis. Thus, the defects of uniform sampling method caused by the number and the location distribution of sample points are overcome and the effect of curve registration is improved. The experimental results on simulated data and real data show that the above two kinds of methods are better than uniform sampling method in time efficiency and the effect of curve registration.
2016 Vol. 29 (1): 72-81 [Abstract] ( 454 ) [HTML 1KB] [ PDF 1159KB] ( 612 )
82 Clustering Algorithm for Mixed Data Based on Dimensional Frequency Dissimilarity and Strongly Connected Fusion
QIAN Chaokai, HUANG Decai
The clustering result of k-Prototypes algorithm is unpredictable due to the sensitivity of the initial prototypes selection. Moreover, the whole diversity between data points and clusters is ignored. Therefore, a clustering algorithm based on dimensional frequency dissimilarity and strongly connected fusion is proposed. Plenty of sub-clusters are produced by multiple pre-clustering. According to the connectivity of those sub-clusters, strongly connected fusion is used to generate the final clusters. The proposed clustering algorithm is validated on three different UCI datasets. Meanwhile, it is compared with three mixed data clustering algorithms. The experimental results show that the proposed algorithm can yield better clustering precision and purity.
2016 Vol. 29 (1): 82-89 [Abstract] ( 428 ) [HTML 1KB] [ PDF 484KB] ( 620 )
90 Label Propagation Algorithm Based on Over-Relaxation Iteration
GE Fang, GUO Youqiang, WANG Nian
Aiming at the problem in the label propagation algorithm, over-relaxation iteration is introduced to solve the optimization problem of label sequence and an improved label propagation algorithm based on over-relaxation iteration (ORLP) is presented. The known samples are labeled with positive and negative labels and the label information of unknown samples is predicted by learning the classification between neighbor points. Meanwhile, the label information of initial labeled samples is reserved in each iteration to guide the next label propagation process. In addition, grounded on over-relaxation iteration, the label propagation formula of ORLP is inferred and the convergence of label sequence is proved simultaneously. Thus, the convergence solution of label sequence is obtained. The experimental results show that the ORLP has higher classification accuracy and convergence speed.
2016 Vol. 29 (1): 90-96 [Abstract] ( 452 ) [HTML 1KB] [ PDF 480KB] ( 856 )
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