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....

2012 Vol.25 Issue.2, Published 2012-04-25 Orignal Article        Orignal Article 181 Dynamic Quotient Space Model and Its Basic Properties ZHANG Ling, ZHANG Bo To solve problems under dynamic conditions, a time variable is introduced based on the original quotient space model (X,f,T), and the original model is extended to a dynamic quotient space (X(t),f(t),T(t)). There are two cases as follows: 1) If structure T is fixed, i.e.,(X(t), f(t),T), the dynamic quotient space model is transformed into a high dimensional static model by introducing a time variable into domain X. Then, the properties of the static model can be used. 2) When both domain X and attribute f are fixed, i.e., (X, f,T(t)), the necessary and sufficient condition for forming a chain of quotient space are discussed. And the corresponding principle of quotient approximation is established and its basic properties are discussed. Finally, the application of the dynamic quotient space model to problem solving is given. 2012 Vol. 25 (2): 181-185 [Abstract] ( 978 ) [HTML 1KB] [ PDF 317KB] ( 730 ) 186 GEP Evolution Algorithm Based on Control of Mixed Diversity Degree XUAN Shi-Bin, LIU Yi-Guang Gene expression programming (GEP) is an evolution algorithm which has the problem of local optimization like other evolution algorithms. The general method to this problem is to keep the diversity degree of population in the evolution. A method is proposed for measuring the diversity of the population, and it merges characters of both population space and sample space. Based on the method for mergence measuring the diversity of population, GEP evolution algorithm with diversity control is proposed. The rival theory is introduced into the initialization of population. The experimental results show that the proposed algorithm efficiently avoids falling into early local optimization. 2012 Vol. 25 (2): 186-194 [Abstract] ( 701 ) [HTML 1KB] [ PDF 578KB] ( 556 ) 195 Image Edge Detection Based on Multi-Granularity Rough Fuzzy Set WANG Dan, WU Meng-Da, MAO Zi-Yang, ZHOU Fan-Cheng An edge detection method based on rough fuzzy set is presented which combines advantages of rough sets that can handle vagueness and C-Set that can provide multi-granularity. An edge expression is proposed by the definition of rough  operation and  operation. Granularity w and structure similarity ε are introduced for better edge detection. The experimental results show that the image edge detection algorithm performs well. 2012 Vol. 25 (2): 195-204 [Abstract] ( 944 ) [HTML 1KB] [ PDF 1600KB] ( 844 ) 205 Cloud Model Based Evolution Strategy Driven by Kurtosis HE Zhen-Feng, XIONG Fan-Lun The Cloud Distribution’s kurtosis statistic and its application are considered to analyze and improve cloud model based evolution strategy (CMES) from the angle of the classical evolution strategy. By adjusting the kurtosis, the Cloud Distribution with a fixed standard deviation changes the shape of noise, which makes the mutation more effective. The formula of the cloud distribution’s kurtosis is derived, which enables the transformation between the entropy-hyper entropy space and the standard deviation-kurtosis space. The influences of the kurtosis and the kurtosis ratio on the cloud distribution noises are compared to prove that the kurtosis is more suitable for self-adaptive. A kurtosis driven CMES, whose parameter evolution combines a 1/5 rule based standard deviation evolution and a self-adaptive kurtosis evolution, is presented. The experimental results of 8 test functions show that a high kurtosis benefits global optimization, a low kurtosis is in favor of local optimization, and the self-adaptive adjustment of the kurtosis can integrate the benefits from both. 2012 Vol. 25 (2): 205-212 [Abstract] ( 602 ) [HTML 1KB] [ PDF 453KB] ( 664 ) 213 WEI Zhen, WU Lei, GE Fang-Zhen, WANG Qiang A hybrid PSO algorithm based on memetic framework (HM-PSO) is proposed. It helps the particles which have certain leaning capacity accelerate convergence rate by Lamarckian Learning based local search strategy and helps the particles which fall into the local optimum escape from local optimum by Tabu search. HM-PSO avoids falling into the local optimum by enhancing the diversity of swarm with accelerating convergence rate. The experimental results show that the improved Lamarckian Learning strategy is effective and feasible and HM-PSO is an effective optimization algorithm with better global search performance. 2012 Vol. 25 (2): 213-219 [Abstract] ( 678 ) [HTML 1KB] [ PDF 443KB] ( 639 ) 220 Closed Frequent Itemset Mining Based on MapReduce CHEN Guang-Peng, YANG Yu-Bin, GAO Yang, SHANG Lin Closed frequent itemset mining is an useful way for discovering association rules from data. Cloud computing infrastructure based on MapReduce provides a promising solution to address the problem. A parallel algorithm for mining closed frequent itemset is presented based on the Hadoop cloud computing platform. The method consists of four steps: parallel counting, global F-List constructing, parallel mining of local closed frequent itemset and parallel filtrating of global closed frequent itemset. The experimental results validate the method and show that it is effective with a satisfied speedup. 2012 Vol. 25 (2): 220-224 [Abstract] ( 1029 ) [HTML 1KB] [ PDF 356KB] ( 776 ) 225 Image Texture Recognition and Retrieval Based on Rough Granular Model XU Jiu-Cheng, LI Xiao-Yan, ZHANG Ling-Jun, LI Shuang-Qun Most of the traditional image texture recognition method focuses on spectrum. In this paper, the layering idea is used based on granular computing theory to recognized the image texture features. Firstly, the rough granular theory is presented and the rough granular space model is constructed by introducing the concept of granular edges and layered entropy. Secondly, a kind of similarity calculation method is established based on granular edges and layered entropy. Then, a method of image texture recognition is put forward. The proposed method improves the practicality of the model and simplifies the calculation of texture recognition by synchronous identification and retrieval. Finally, the simulation result show that the performance of image retrieval is improved. The validity of the proposed method is testified by the contrast with other methods. 2012 Vol. 25 (2): 225-229 [Abstract] ( 620 ) [HTML 1KB] [ PDF 287KB] ( 720 ) 230 Video Copy Detection Based on Spatio-Temporal Feature ZHANG Zhi-Jie, ZOU Jian-Hua A video copy detection method based on affine invariant and compact spatio-temporal feature is proposed. The proposed method is based on the maximally stable extremal blocks (MSEB), and the use of the MSEB is more in conformity with characteristics of the human visual feeling perception than that of local features. Firstly, several continuous frames of a video are piled into a 3-D video cube. Then, the MSEB in each cube are detected based on the watershed thoughts. After the detection of MSEB, each MESB is described by a serial of 3-D moment invariants. Finally, copies of videos are accurately identified by the MSEB. The experimental results show that the proposed method is effective in matching. 2012 Vol. 25 (2): 230-236 [Abstract] ( 635 ) [HTML 1KB] [ PDF 426KB] ( 498 ) 237 Total Margin Based Fuzzy Hypersphere Learning Machine TAO Jian-Wen, WANG Shi-Tong There are several problems in classical support vector machines, such as overfitting problem resulted from the outlier and class imbalance learning and the loss of the statistics information of training examples. Aiming at these problems, a total margin based fuzzy hypersphere learning machine (TMF-SSLM) is proposed by constructing a minimum hypersphere in Mercer kernel-induced feature space. The main idea of TMF-SSLM is that one class of binary patterns is enclosed in the minimum hypersphere, from which another one is separated away with maximum margin. Thus both maximum between-class margin and minimum within-class volume are implemented. The proposed TMF-SSLM solves the overfitting problem resulted from outliers by employing both the fuzzification of the penalty and total margin algorithm, as well as the imbalanced problem by using different cost algorithm. Theoretical analysis justifies that TMF-SSLM obtains a lower generalization error bound. The exprimental results obtained on real datasets show that the proposed algorithm is stable and superior to other related diagrams. 2012 Vol. 25 (2): 237-247 [Abstract] ( 758 ) [HTML 1KB] [ PDF 692KB] ( 690 ) 248 Structured Gait Feature Expression and Fast Gait Recognition Method WEI Su-Yuan, NING Chao, GAO You-Xing , LI Gang When the gait database contains the only gait feature, the larger the number of individuals is, the longer time the recognition algorithm costs and the lower the recognition ratio is. Aiming at this problem, a structured gait feature expression and fast gait recognition method is presented. The structured gait feature are made up of gait information, individual height, sex, age and other information. These feature components are sampled by different sensors and used independently. The proposed gait recognition algorithm utilizes the structured feature to deal with the gait recognition hierarchically. The large identification range is narrowed. The experimental results demonstrate that the proposed method improves the recognition speed and gains higher identification precision. 2012 Vol. 25 (2): 248-255 [Abstract] ( 642 ) [HTML 1KB] [ PDF 561KB] ( 624 ) 256 Bi-L1 Sparse Representation Algorithm for Face Recognition Based on Fusion of Global and Separated Components HU Zheng-Ping, SONG Shu-Fen Considering the complementation of global and local components, bi-L1 sparse representation algorithm for face recognition based on fusion of global and separated components is proposed. Firstly, based on L1 sparse representation, the global information is used to obtain the global sparse approximation. Then, several slightly overlapping face components are extracted and aligned in the recognition model of separated components. After that, the sparse representation of all the components is obtained respectively. The sparse approximation results of each component are combined with a similarity voting method based on the residuals of class representation. Finally, the weighted integration of the global and components sparse representation is used to construct the bi-L1 sparse representation classifier in decision-making layer. The experimental results on public available database demonstrate that the performance of the integration classifier is superior to that of each single module. Due to the fusion of component information which is insensitive to variation of illumination and expression etc., the robustness of the system is enhanced. 2012 Vol. 25 (2): 256-261 [Abstract] ( 311 ) [HTML 1KB] [ PDF 406KB] ( 642 ) 262 Task Allocation in Networked Multiagent Systems JIANG Yi-Chuan Large scale multiagent systems are always organized in networked structures where each agent interacts only with its immediate neighbors. Moreover, the networked multiagent systems always run on certain underlying physical networks. Obviously, the traditional task allocation methods based on agent self-owned resources are not fit for the networked multiagent systems. Aiming at this problem, three task allocation methods are reviewed for networked multiagent systems: the task allocation method based on underlying networks and agent resources, the task allocation method based on multiagent interaction networks and agent resources and the task allocation method based on contextual resource distribution. It considers both the underlying networks and multiagent interaction networks. Besides, the related works on centralized and distributed task allocations are reviewed, and the related works are compared to the proposed task allocation methods. Finally, the difficulties and the further work on the task allocation of networked multiagent systems are discussed. 2012 Vol. 25 (2): 262-272 [Abstract] ( 615 ) [HTML 1KB] [ PDF 806KB] ( 714 ) 273 Time Series Analysis Based Human Fall Prediction Method TONG Li-Na ,SONG Quan-Jun ,GE Yun-Jian A method for human fall prediction based on time series of human action states is proposed. Firstly, the acceleration time series in characteristic body region is got by information fusion procedure. Secondly, the segments before the collision of body with lower objects in fall processes is chosen as samples to train hidden Markov model (HMM). Then, the current-time fall risk is analyzed by the real-time matching degree between input series and HMM. The experimental result shows that the proposed method gets good result in predicting falls, and the fall events and other daily life activities can be distinguished effectively by it. 2012 Vol. 25 (2): 273-279 [Abstract] ( 856 ) [HTML 1KB] [ PDF 520KB] ( 1070 ) 280 Digital Image Forensics Based on Local Energy Variance Properties QIAO Tong, QIAN Zhen-Xing, ZHANG Xin-Peng, WANG Wen-Wen How to identify photographic images and computer graphics efficiently is more and more important in image forensics. An authentication algorithm for digital images is proposed using the correlations between photographic images and computer graphics. The calculated local energy dithering in the frequency domain is used to distinguish photographic images and computer-generated images. Moreover, compared with Gallagher’s method, the proposed method is still available when both kinds of images above are attacked by CFA interpolation or zooming operation. The experimental results show that the proposed algorithm identifies better with the capability of anti-forensics resistance. 2012 Vol. 25 (2): 280-284 [Abstract] ( 583 ) [HTML 1KB] [ PDF 343KB] ( 774 ) 285 Combined-Feature-Discriminability Enhanced Canonical Correlation Analysis ZHOU Xu-Dong, CHEN Xiao-Hong, CHEN Song-Can Canonical Correlation Analysis (CCA) has following two deficiencies in performing classification task: CCA can not directly optimize them but their components, though combined features are the input of the classifier; CCA can not utilize any class information at all, though facing classification task. To overcome these deficiencies, a supervised dimension reduction method named combined-feature-discriminability enhanced canonical correlation analysis (CECCA) is proposed. CECCA is developed through incorporating discriminant analysis of combined features into CCA. Consequently, it optimizes the combined feature correlation and discriminability simultaneously and thus makes the extracted features more suitable for classification. The experimental results on artificial dataset, multiple feature database and facial databases show that the proposed method is effective. 2012 Vol. 25 (2): 285-291 [Abstract] ( 778 ) [HTML 1KB] [ PDF 433KB] ( 853 ) 292 Semi-Supervised Learning Based Ensemble Classifier for Stream Data XU Wen-Hua , QIN Zheng , CHANG Yang Stream data classification algorithms are mainly based on supervised learning strategy, and they need massive labeled data for training. These approaches are unpractical due to the high cost of acquiring labeled data in a real streaming environment. A semi-supervised learning based ensemble classifier (SEClass) is presented for stream data classification. SEClass utilizes both a small number of labeled data and a great number of unlabeled data to train an ensemble classifier, and unlabeled instances are classified using the majority voting strategy. The experimental results show that the accuracy of SEClass is 5.33% higher in average than that of the state-of-the-art supervised method using the same number of labeled data for training. And SEClass is suitable for high-dimensional high-speed massive stream data classification. 2012 Vol. 25 (2): 292-299 [Abstract] ( 817 ) [HTML 1KB] [ PDF 563KB] ( 821 ) 300 A Multi-Resolution Medical Image Fusion Method Based on Nonparametric Orthogonal Polynomials LIU Zhe, SONG Yu-Qing, CHEN Jian-Mei, YAN Yan-Hua, XIE Cong-Hua Image fusion algorithms based on estimation theory assume that all distortions should follow Gaussian distribution, which causes model mismatch, local details losing and time-consuming. A medical image fusion algorithm is presented based on multi-resolution and nonparametric orthogonal polynomials. Firstly, the source images are decomposed into multi-resolution representations. Then, the NEM algorithm is used to estimate the parameters of the image information model and the non-parametric orthogonal polynomials image mixture model. Thus, the fusion result of low frequency band image is got. For the high frequency band image, the maximum absolute value of the coefficient is applied. Finally, the fused image is obtained by taking the inverse transformation of the composed coefficients. The experimental results show that the proposed algorithm achieves better performance than other fusion methods and the fusion time is considerably reduced. 2012 Vol. 25 (2): 300-304 [Abstract] ( 527 ) [HTML 1KB] [ PDF 626KB] ( 599 ) 305 Discriminative Feature Fusion Based on Extensions of PCA TAN Jing-Dong, SU Ya-Ru, WANG Ru-Jing Two methods for dimensionality reduction, principal component discriminative analysis and kernel principal component discriminative analysis, are proposed. Based on the theory of principal component analysis and maximum margin criterion, a multi-objective project model is constructed to formalize the goals for feature fusion. Then, it is transformed into a single-objective cost function for the projection, and the optimal linear mapping is obtained through optimizing this cost function. Additionally, the nearly diagonal block kernel matrix is divided into c kernel matrixes (c is the number of classes in dataset), and eigen-decomposition method is used to solve their d principal vectors. Through the process of vector algebra, a combined mapping α is obtained. When the original kernel matrix K is projected on α, the inner-class information is optimally preserved. The experimental results show their validity. 2012 Vol. 25 (2): 305-312 [Abstract] ( 669 ) [HTML 1KB] [ PDF 518KB] ( 728 ) 313 Color Image Retrieval Using Zernike Chromaticity Distribution Moments WANG Xiang-Yang, LI Dong-Ming, YANG Hong-Ying Feature extraction and representation are important to the content-based image retrieval (CBIR). Color is one of the widely used visual features and it is invariant to image size and orientation. According to the opponent chromaticity space and Zernike moments, a color image retrieval using Zernike chromaticity distribution moments is proposed. Firstly, the image is transformed from RGB space to opponent chromaticity space. Then, the characteristics of the color contents of an image are captured by using Zernike chromaticity distribution moments directly from the chromaticity space. Finally, the similarity between color images is computed by Zernike chromaticity distribution moments. The experimental results show the effectiveness of the proposed algorithm in improving the retrieval performance. 2012 Vol. 25 (2): 313-317 [Abstract] ( 643 ) [HTML 1KB] [ PDF 343KB] ( 1198 ) 318 Mahalanobis Distance Measurement Based Locally Linear Embedding Algorithm ZHANG Xing-Fu, HUANG Shao-Bin Euclidean distance is normally used to measure the similarity between samples in locally linear embedding algorithm(LLE). But for some high dimensional data, such as images, Euclidean distance can not accurately reflect the similarity between samples. A Mahalanobis distance metric based locally linear embedding algorithm (MLLE) is proposed. Firstly, MLLE ascertains a Mahalanobis metric from the existing samples. Then, the Mahalanobis metric is used to choose neighborhoods and to reduce the dimensionality of the existing samples and the new samples. The comparison result of MLLE algorithm and some classical manifold based algorithms on ORL and USPS databases proves that MLLE algorithm is effective in recognizing images. 2012 Vol. 25 (2): 318-324 [Abstract] ( 771 ) [HTML 1KB] [ PDF 415KB] ( 618 ) 325 A Text Region Location Method Based on Connected Component YAO Jin-Liang, WENG Lu-Bin, WANG Xiao-Hua Text region location is important to text recognition and retrieval in images of complex background. The existing methods with precision and recall rate have high computational complexity. These methods are unpractical real environment. A text region location method is proposed based on component filtering and K-means clustering. Firstly, the input image is segmented into three layers by an adaptive image segmentation method, and the components are extracted from the character layers. Then, the features of the component are obtained, and Adaboost classifier is used to filter non-character components. The candidates of character components are grouped into text regions by K-means clustering based on the position and layer of the component. The experimental results demonstrate that the precision and the recall rate of the proposed approach is almost the same that of as the other methods, and the proposed method has lower computational complexity. 2012 Vol. 25 (2): 325-331 [Abstract] ( 640 ) [HTML 1KB] [ PDF 797KB] ( 948 ) 332 Flexible Feature Optimization Based Robust Object Tracking WANG Jiang-Tao, CHEN De-Bao, YANG Jing-Yu To overcome the disadvantages of single feature that often fails in describing the object reliably under dynamic environment, a flexible feature optimization based particle filter tracking algorithm is proposed. Firstly, the concept of sharpness factor is introduced to objectively elevate the discriminant ability for different features. Then, based on the features tracking property, the optimal feature under current scene is adaptively generated by combining the weighted features. Finally, the optimal feature is applied in the particle filter scheme to execute the object tracking task. The proposed algorithm is flexible and it can be extended to any feature represented by histogram. The experimental results on various videos demonstrate the effectiveness and robustness of the proposed method in multi-features fusion and object tracking. 2012 Vol. 25 (2): 332-338 [Abstract] ( 601 ) [HTML 1KB] [ PDF 1085KB] ( 799 ) 339 Text Classification Using Diffusion Kernel on Statistical Manifold LI Kan, ZHOU Shi-Bin, LIU Yu-Shu Dirichlet compound multinomial manifold (DCM manifold) is proposed. DCM manifold with positive sphere manifold is homeomorphic and isometric, so the geodesic distance of positive sphere manifold can be mapped as the geodesic distance of DCM manifold through pullback mapping. Then the distance metric is built on DCM manifold. DCM diffusion kernel function and DCMIDF diffusion kernel function are built on DCM manifold. The performance of the proposed algorithms for text classification are tested on the corpuses of WebKB Top 4 and 20 Newsgroups, and the experimental results show that DCM manifold is more desirable than that of Euclidean space in modeling texts on the corpuses. Compared with polynomial kernel based support vector machine and NGD kernel based support vector machine, the proposed DCM diffusion kernel and DCMIDF diffusion kernel based support vector machine algorithms show better computational accuracy for text classification. 2012 Vol. 25 (2): 339-345 [Abstract] ( 635 ) [HTML 1KB] [ PDF 459KB] ( 842 ) 346 Fusion Recognition Algorithm Based on Fuzzy Density Determination with Classification Capability and Supportability ZHAN Yong-Zhao, ZHANG Juan, MAO Qi-Rong Fuzzy integral theory can be effectively used to deal with the uncertainties of the classification decisions. However, the classification capability of each classifier for recognition results and the supportability of each classifier for the object recognition are not taken into account in the current methods of fuzzy density determination, which results in the loss of the important information for fusion recognition. To overcome this disadvantage, a fusion recognition algorithm based on fuzzy density determination with classification capability and supportability for each classifier is presented. In this algorithm, the fuzzy densities for the classifier fusion are adaptively determined by classification capability of each classifier for recognition results and supportability of each classifier for the object recognition. Thus, the multi-classifiers fusion recognition can be effectively realized. The proposed algorithm is used to recognize facial expression in natural interaction situation and Cohn-Kanade facial expression database. The experimental results show that the proposed algorithm effectively raises the accuracy of expression recognition. 2012 Vol. 25 (2): 346-351 [Abstract] ( 600 ) [HTML 1KB] [ PDF 441KB] ( 667 ) 352 Human Motion Data Retrieval Based on Dynamic Time Warping Optimization Algorithm LIU Xian-Mei, ZHAO Dan, HAO Ai-Min With the emergence of many large-scale three-dimensional human motion databases, content-based retrieval of 3D human motion faces many difficulties. A human motion data retrieval technology based on improved dynamic time warping optimization algorithm is proposed, by which logically similar motions can be found effectively. Firstly, the coordinates of two motion sequences are aligned and a distance matrix based on window of frames is constructed. Then, using an optimization algorithm based on global and local constraints, a similarity matching is processed to describe corresponding relationship between two motions. Finally, similar motions are retrieved by two-phase approach with normalization similarity and DTW average distance. The experimental results show that by two-phase DTW optimization approach, better retrieval results for motions which are not aligned in time axis are obtained and the efficiency is improved. 2012 Vol. 25 (2): 352-360 [Abstract] ( 577 ) [HTML 1KB] [ PDF 616KB] ( 827 )

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