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....

2015 Vol.28 Issue.12, Published 2015-12-30 Papers and Reports    Researches and Applications        Papers and Reports 1057 Descriptive Method on Data Association Based on Granulation Trees YAN Shuo , YAN Lin To study the data association, a data set is divided into different hierarchy granules. Consequently, a hierarchy structure called a granulation tree is obtained. Then, grounded on the granular hierarchy information in the granulation tree, the numerical representation of granularity and the data connections determined by association data, a definition of the data association between two granulation trees is introduced. In this paper, the upper approximation is taken as an operator. With the granule corresponding to the operation of the upper approximation, a theorem is established, and it is used to investigate whether two data are associated with each other. The investigation bases the close degree of the association on the numerical information of granular. Therefore, a method taking granulation trees as the basis to describe the data association is presented. The characteristics of the hierarchy granules and the numerical representation of granularity contained in the method provide a way of research on granular computing. The discussion on a specific example shows the application value of the granulation tree method. 2015 Vol. 28 (12): 1057-1066 [Abstract] ( 392 ) [HTML 1KB] [ PDF 402KB] ( 453 ) 1067 Moment-Optimized Boosting Algorithm LIU Chuan, LIAO Shi-Zhong Margin distribution is critical to Boosting. However, the existing margin-based generalization error bounds are too complicated to be used for the design of new Boosting algorithms. In this paper, a moment-optimized Boosting (MOBoost) algorithm is proposed with direct optimization of the margin distribution. Firstly, a generalization error bound for Boosting based on first and secondary moments of the margin distribution is derived to reveal the close relationship between margin distribution and generalization error. Then, a moment criterion for Boosting model selection is presented based on the moment generalization bound. The criterion maximizes the first moment and minimizes the second moment of the margin distribution simultaneously. Consequently, the primary and dual forms are formulated for solving the convex quadratic program of the moment criterion for Boosting. Thus, an efficient computing method for the moment criterion is proposed. Theoretical analysis and experimental results show that MOBoost is effective and reliable. 2015 Vol. 28 (12): 1067-1073 [Abstract] ( 399 ) [HTML 1KB] [ PDF 482KB] ( 527 ) 1074 Improved RDS Symbol Algebraic Decision Diagram Algorithm for Weighted Constraint Satisfaction Problem XU Zhou-Bo, YANG Xin-Liang, GU Tian-Long, Ning Li-Hua Weighted constraint satisfaction problem (WCSP) is a kind of constraint optimization problem. Based on Russian doll search (RDS) algorithm, an improved RDS symbolic algebraic decision diagram (ADD) algorithm for WCSP is proposed to reduce the number of sub-problems and improve the efficiency of solving sub-problems of original RDS algorithm. Through improving the most constraining variable (MCV) method of variable selection, a concept of RDS variable is introduced and conducted for nested decomposition of the original problem. Then, the number of sub-problems is decreased. Furthermore, the nested decomposition method is improved by variable back-degree. To improve the efficiency of solving sub-problems, the bucket elimination algorithm combined with ADD operation is exploited to eliminate the non-RDS variables. And thus the number of variables in the sub-problem is decreased and the lower bound is improved. Experiments on a large number of random generated testing cases demonstrate the superiority of the proposed algorithm. 2015 Vol. 28 (12): 1074-1083 [Abstract] ( 334 ) [HTML 1KB] [ PDF 451KB] ( 527 ) 1084 A Differentially Private Histogram Publication Algorithm for Arbitrary Range Tree Structure WU Ying-Jie, CHEN Hong, WANG Yi-Lei, SUN Lan In the existing methods for differentially private histogram publication, a histogram is mapped to a perfect m-ary range tree. The accuracies of queries are boosted through consistency constraints of the queries. However, not all histograms in real application can be mapped to perfect m-ary range trees directly. In this paper, a range tree structure, k-range tree, is firstly put forward. By k-range tree, an arbitrary histogram is mapped to a range tree. Secondly, the theoretical analysis shows that for differentially private histogram publication for arbitrary range tree structure, the error of range counting queries still can be further reduced by solving the best linear unbiased estimation of the tree node values through consistency. Finally, a differentially private histogram publication algorithm based on local best linear unbiased estimation(LBLUE) for arbitrary range tree structure is proposed. Experiment is carried out to compare LBLUE and the traditional algorithms on the accuracy of range counting queries in the released histogram and the algorithm efficiency. Experimental results show that LBLUE is effective and feasible. 2015 Vol. 28 (12): 1084-1092 [Abstract] ( 496 ) [HTML 1KB] [ PDF 523KB] ( 682 ) Researches and Applications 1093 Language Identification Based on Deep Neural Network CUI Rui-Lian, SONG Yan, JIANG Bing, DAI Li-Rong Aiming at the problems of confusable dialects and short-duration utterance in automatic spoken language identification (LID), an improved utterance representation method is proposed based on different layers of deep neural network (DNN). Deep bottleneck network (DBN), a DNN with an internal bottleneck layer, is employed as a front-end feature extractor. Different representations based on output layer and middle bottleneck layer of DBN for LID are obtained and fused. Evaluations on the NIST LRE2009 dataset and NIST LRE2011 Arabic dialect dataset demonstrate that the proposed method based on DBN achieves good performance. 2015 Vol. 28 (12): 1093-1099 [Abstract] ( 607 ) [HTML 1KB] [ PDF 552KB] ( 1841 ) 1100 Center-Based Line Neighborhood Discriminant Embedding Algorithm and Its Application to Face Recognition YANG Zhang-Jing , HUANG Pu , ZHANG Fan-Long , YANG Guo-Wei To overcome the drawbacks of the existing marginal fisher analysis algorithm in feature extraction, a center-based line neighborhood discriminant embedding (CLNDE) algorithm is proposed for face recognition. Firstly, the distance from a sample point to the center-based line is utilized to construct the within-class similarity matrix and the between-class similarity matrix, respectively. Next, the between-class local scatter and the within-class local scatter of samples are calculated by the constructed similarity matrices. Finally, the optimal transformation matrix is found by maximizing the between-class local scatter and minimizing the within-class local scatter simultaneously. Experimental results on face databases demonstrate the superiority of the proposed algorithm. 2015 Vol. 28 (12): 1100-1109 [Abstract] ( 530 ) [HTML 1KB] [ PDF 669KB] ( 480 ) 1110 JavaScript Malicious Script Detection Algorithm Based on Multi-class Features FU Lei-Peng, ZHANG Han, HUO Lu-Yang Aiming at features of different levels in the script sample set, such as obfuscation, statistics and semantics, a malicious JavaScript script detection algorithm based on multi-class feature is proposed. The JavaScript analysis system, JavaScript codes analysis and detection, is implemented. The obfuscation features of the JavaScript are extracted and the obfuscated scripts are analyzed and deobfuscated by C4.5 algorithm. The static statistical features of the JavaScript are extracted, and according to the semantics, the JavaScript is serialized. Dangerous sequence tree is generated by the proposed algorithm to extract the dangerous sequence features of the malicious JavaScript. Three types of features are used as the input. The probabilistic neural network with strong ability to adapt to non-uniformity and the increasing quantity of the input samples is applied to construct the classifier for the detection of malicious JavaScript. The experimental results show that the proposed algorithmhas better detection accuracy and stability. 2015 Vol. 28 (12): 1110-1118 [Abstract] ( 464 ) [HTML 1KB] [ PDF 544KB] ( 1213 ) 1119 Three-Dimensional Model Similarity Analysis Based on Salient Features Spectral Embedding HAN Li , YAN Zhen , XU Jian-Guo , TANG Di Aiming at the requirement of efficient 3D model retrieval technology, a three-dimensional model similarity analysis based on salient features spectral embedding method is proposed. Firstly, the salient features are extracted by curvature-based method and a convex-concave measurement to build the salient features representation for the shape. Then these features are embedded in a spectral domain to reveal the intrinsic shape characteristics based on Laplacian Eigenmap. Finally, combined with the thin plate splines method, the model similarity analysis and registration are implemented. The experimental results show that by using the proposed method shape matching is implemented efficiently and the consistent structural features in same category models are identified. Moreover, it is robust to the imperfect shape matching. 2015 Vol. 28 (12): 1119-1126 [Abstract] ( 440 ) [HTML 1KB] [ PDF 905KB] ( 821 ) 1127 Human Activity Recognition Based on Acceleration Signal and Evolutionary RBF Neural Network LU Xian-Ling, WANG Hong-Bin, XU Xian To obtain a satisfactory recognition rate,a radial basis function(RBF) neural network classifier trained by the hierarchy genetic algorithm (HGA) is utilized to classify human body activities using the acceleration signal. By exploring the interquartile range, a fitness function is proposed to enhance the crossover of the parameter genes in HGA and determine the distance between the offspring and the boundary of coding space automatically. Thus, the empirical setting in the previous algorithms is avoided. With the arithmetic crossover, the offspring with high fitness is chosen. By comparing fitness values between the uniform mutation offspring and the non-uniform mutation offspring, the structure and parameters of RBF network are jointly optimized. The experimental results on actual subject testing data indicate that the radial basis function neural network classifier trained by the proposed method produces smaller errors than those trained by the traditional HGA. A higher recognition rate of testing data is obtained. 2015 Vol. 28 (12): 1127-1136 [Abstract] ( 557 ) [HTML 1KB] [ PDF 1319KB] ( 504 ) 1137 Stock Market Trend Forecast Algorithm Based on Energy Computational Model of Bayesian Networks ZHANG Run-Mei , HU Xue-Gang , WANG Hao , YAO Hong-Liang There are inconsistencies between the technical indexes and the trend of stock market, and therefore the trend of stock market is difficult to predict. Through the energy characteristics extraction and the feature fusion of technical indexes, a stock market trend forecast algorithm based on energy calculation of Bayesian networks (E-STF) is proposed. Firstly, the trend information inside technical indexes is extracted from the point of energy, the energy calculation model of technical indexes is designed and its probability distribution is given. Then, the inconsistency of energy distribution between technical indexes is analyzed. Next, Bayesian networks are used to fuse the features of technical indexes. The time-sharing state feature energy is introduced and fused with technical indexes energy to build the stock market trend structure model. Finally, based on the conditional probability function between the stock market trend and some relative characteristic energy, energy constraints relationship is introduced into support vector machine to predict the stock market trend. Through comparison and analysis on the Shanghai Stock Exchange indexes in recent 3 years, the experimental results show that the prediction accuracy is improved effectively by E-STF algorithm. 2015 Vol. 28 (12): 1137-1146 [Abstract] ( 657 ) [HTML 1KB] [ PDF 618KB] ( 1111 )

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