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....

2010 Vol.23 Issue.4, Published 2010-08-31 Orignal Article        Orignal Article 441 Ensemble Classifier Based on Minimum Class Variance SVM and Null Space Classifier WANG Xiao-Ming, WANG Shi-Tong The Minimum Class Variance Support Vector Machine (MCVSVM) takes into consideration both the samples in the boundaries and the distribution of the classes. However, only the information in the non-null space of the within-class scatter matrix is utilized in the case of small sample size. To further improve the classification performance, in this paper the Null Space Classifier (NSC) which is rooted in the null space is first presented, then an Ensemble Classifier (EC) is proposed by fusing the MCVSVM and the NSC. Different form the MCVSVM and the NSC, the EC considers the information both in the non-null space and in the null space and has better generalizability. Finally, experimental results on several real datasets indicate the effectiveness of the EC. 2010 Vol. 23 (4): 441-449 [Abstract] ( 296 ) [HTML 1KB] [ PDF 556KB] ( 489 ) 450 A Constrained Line Search Optimization Method for Discriminative Training of HMMs LIU Cong,HU Yu,DAI Li-Rong,WANG Ren-Hua In this paper a optimization algorithm called constrained line search (CLS) is proposed for discriminative training (DT) of Gaussian mixture Continuous Density Hidden Markov Model (CDHMM) in speech recognition. The CLS method is used to optimize the objective function of Maximum Mutual Information (MMI) criterion based discriminative training. In this method, discriminative training of HMM is firstly treated as a constrained optimization problem, and Kullback-Leibler divergence (KLD) between models is explicitly applied as a constraint during optimization. Based upon the idea of line search, it is pointed out that a simple formula of HMM parameters can be expressed as a simple quadratic formula by constraining the KLD between HMM of two successive iterations. The proposed CLS method can be applied to optimize all model parameters in Gaussian mixture CDHMMs, including means, covariances, and mixture weights. The proposed CLS approach is investigated on two benchmark speech recognition tasks of both English and Chinese, including TIDIGITS and Mandarin 863 database. Experimental results show that the CLS optimization method consistently outperforms the conventional extended Baum-Welch (EBW) method in both recognition performance and convergence characteristic. 2010 Vol. 23 (4): 450-455 [Abstract] ( 300 ) [HTML 1KB] [ PDF 442KB] ( 718 ) 456 Hierarchical Text Classification Model Based on Blocking Priori Knowledge LI Wen,MIAO Duo-Qian,WEI Zhi-Hua,WANG Wei-Li Blocking exerts negative effect on the performance of text hierarchical classification. In this paper, a two-step hierarchical text classification model based on blocking priori knowledge is proposed to address the problem. Firstly, blocking distribution is estimated and blocking pair recognition technique focusing on mining the serious blocking direction is presented. Secondly, the hierarchy topology structure is actively refined which attempts to correct misclassification and reduce blocking errors by using blocking priori knowledge. The experimental results on TanCorp, which is a new corpus special for Chinese text classification, show that the model can improve the performance significantly without increasing the extra number of classifiers and is a method of solving the hierarchical classification blocking problem. In addition, compared with flat text classification algorithm, this method has stable performance. 2010 Vol. 23 (4): 456-463 [Abstract] ( 352 ) [HTML 1KB] [ PDF 516KB] ( 451 ) 464 QR Code Based Semantic Map Building in Domestic Semi-Unknown Environment WU Hao,TIAN Guo-Hui,XUE Ying-Hua,ZHANG Tao-Tao Aiming at robot special service tasks and man-machine interaction in a home environment, a semantic room map of semi-unknown environment is built using QR code based self-similar artificial object mark plastered on large objects. First, a topology map with the function of room segmentation is built based on spectral clustering algorithms. Then object information database and adscription relationship map are set up based on object information stored in the QR code. Finally, a semantic map including object information description, room functional information and attributive relation between room and object is formed, which gives complete and anthropopathic information for object location, object management and robotic service in a home environment. The simulation results show that the service robot using semantic map can understand human semantic statement, produce reasonable service path and achieve function-driving navigation. 2010 Vol. 23 (4): 464-470 [Abstract] ( 344 ) [HTML 1KB] [ PDF 518KB] ( 707 ) 471 A Competition Optimization Mode of Ecological Particle Swarm AN Jing,KANG Qi,WANG Lei,WU Qi-Di An ecological particle swarm competition optimization model is proposed in this paper by introducting the original idea of population density in ecology into swarm intelligent computation. The dynamics characteristics can more fully describe individuals, environment and cooperative behavior between them, which is to a certain extent out of the biological evolution framework only applying individual fitness to control the evolution. Numerical simulation results show that the proposed ecological PSO model can effectively improve the premature convergence and convergence speed. 2010 Vol. 23 (4): 471-476 [Abstract] ( 306 ) [HTML 1KB] [ PDF 428KB] ( 543 ) 477 Face Recognition Based on Enhanced Gabor Feature and Direct Fractional-Step Linear Discriminant Analysis ZOU Jian-Fa, WANG Guo-Yin, GONG Xun Gabor features can effectively represent the local features of face image with different directions and scales. However, traditional Gabor features based algorithms neglect the global features of the original image. Enhanced Gabor features (EGF) is developed in this paper by combining Gabor features and information extracted from the original image. A face recognition method is further proposed based on EGF and direct fractional-step linear discriminant analysis algorithm (DF_LDA). Experiment results of simulation on Yale, ORL and Georgia face databases show that EGF can effectively improve the face recognition rate compared with the traditional Gabor features. 2010 Vol. 23 (4): 477-482 [Abstract] ( 337 ) [HTML 1KB] [ PDF 425KB] ( 892 ) 483 High-Dimensional Indexing Method Based on Elliptical-Shaped Clustering CUI Jiang-Tao, GUO Yong, ZHOU Shui-Sheng A high-dimensional linear indexing method is presented by sorting principal component based on elliptical-shaped clustering. The proposed approach reduces the number of data points accessed during the k-nearest neighbor search. The dataset is partitioned into some elliptical-shaped clusters, and KL transform is performed on each cluster. The approximate vectors are built at the KL transform domain on each cluster. When performing k-nearest neighbor search, the partial distortion searching algorithm is used to reject the improper approximate vectors. The clusters are accessed in increasing order of their lower bound from the query point. The experimental results on large image databases with high dimensions show that compared with other well-known vector approximate method, the proposed approach reduces the number of approximate vectors accessed and provides a higher search speed. 2010 Vol. 23 (4): 483-490 [Abstract] ( 285 ) [HTML 1KB] [ PDF 559KB] ( 497 ) 491 SVM Incremental Learning Using Simulated Cutting Algorithm SHEN Feng-Shan,ZHANG Jun-Ying ,WANG Kai-Jun A method named Simulated Cutting Algorithm (SCA) is introduced for SVM incremental learning. SCA computes the anticipated contribution for the mapped target of each training sample in feature space mapped by a kernel function, and then chooses samples with higher anticipated contribution for SVM incremental learning. It effectively solves the problems in traditional incremental learning, such as higher training cost, lower accuracy for selecting target samples and lacking robustness. The anticipated contribution rate of a sample is indicated by the recognition rate towards two classes of samples of an appropriate separating hyperplane going through the mapped target of this sample point. Since the way for choosing target samples is very similar to that for paring garden stuff, the proposed algorithm acquires its name from this. Numerical experiments on benchmark datasets show the proposed method is superior in learning efficiency and generalization performance of a classifier. The application of the proposed algorithm in learning with limited resources demonstrates its excellent performance in large-scale learning tasks. 2010 Vol. 23 (4): 491-500 [Abstract] ( 327 ) [HTML 1KB] [ PDF 683KB] ( 585 ) 501 Adaptive Maximal Rejection Discriminant Analysis and Its Discriminant Vectors GUO Zhi-Bo,YAN Yun-Yang,YANG Jing-Yu,ZHAO Chun-Xia Aiming at the two-value output of weak classifiers and Non-orthogonal discriminant vectors on the MRC-Boosting, an adaptive maximal rejection discriminant analysis (AdaMRDA) is proposed to further improve the classification performance. Based on the Euclid distance between the extracted discriminant features and their expectation mean, an adaptive updating weights method is developed firstly, by which the latter discriminant vectors obtained are more beneficial for classification. Then the equation solving the optimal orthogonal discriminant vectors is given. Finally, the experimental results on 2 databases prove that AdaMRDA is superior to MRC-Boosting and related methods on classification performance. 2010 Vol. 23 (4): 501-507 [Abstract] ( 290 ) [HTML 1KB] [ PDF 451KB] ( 534 ) 508 A Smoothed Boosting Algorithm for Ensemble Parameters Learning of Bayesian Network Classifiers WANG Zhong-Feng,WANG Zhi-Hai,FU Bin The property of sample confidence measure function applied by ensemble algorithm of reducing noises is firstly analysed in this paper, and the reason of this function being unfit for multiclass dataset is expounded. Then a confidence measure function with more pertinence is designed, and an enhanced algorithm for reducing noises and ensemble parameters is proposed based on this function. Thus the discriminative parameters learning algorithm of Bayesian network not only effectively restrains the noise impact, but also avoids over fitting of classifiers, and further extend the application of discriminative Bayesian network calssifier applying ensemble learning algorithm in multiclass problem. Finally, the experimental results and its analysis on statistical hypothesis test verify that this algorithm more notably improves the classifier performance than ensemble parameters learning algorithms of Bayesian network at present. 2010 Vol. 23 (4): 508-515 [Abstract] ( 297 ) [HTML 1KB] [ PDF 507KB] ( 656 ) 516 An Efficient Method for K-Means Clustering HUANG Zhen-Hua, XIANG Yang, ZHANG Bo, WANG Dong, LIU Xiao-Ling The existing K-Means clustering methods directly act on multidimensional datasets. Hence, these methods are extremely inefficient as the cardinality of input data and the number of clustering attributes increase. Motivated by the above fact, in this paper, an efficient approach for K-Means clustering based on the structure of regular grid, called KMCRG (K-Means Clustering based on Regular Grid), is proposed. This method effectively implements K-Means clustering by taking cell as handling object. Especially, this method uses the tactics of grid weighted iteration to effectively gain the final K classes. The experiment results show that the algorithm can quickly gain the clustering results without losing clustering precision. 2010 Vol. 23 (4): 516-521 [Abstract] ( 300 ) [HTML 1KB] [ PDF 391KB] ( 663 ) 522 Non-Standard Inferences in Description Logics TANG Su-Qin,CAI Zi-Xing ,WANG Ju,JIANG Yun-Cheng Description Logics belong to a kind of formalization method for representing knowledge in the knowledge engineering. Solving the standard and non-standard inferences has become an important research area of description logics in recent years. The significance and research advance on standard and non-standard inferences in description logics are summarized in this paper. The definition of Least Common Subsumer(LCS), Most Specific Concept(MSC), Rewriting, Matching, Debugging and Conservative Extensions, and the actualizing technologies of these Reasoning are given. Problems of the LCS, MSC, Matching, non-standard reasoning of hybrid terminological cycles in description logics and their research advance are discussed in particular. Finally, some prospects of non-standard inference in description logic are discussed. 2010 Vol. 23 (4): 522-530 [Abstract] ( 324 ) [HTML 1KB] [ PDF 753KB] ( 581 ) 531 Rough Set Models Based on Fuzzy Inclusion and Fuzzy Belief Measures ZHANG Jia-Lu,ZHAO Xiao-Dong In this paper, a fuzzy rough set model based on fuzzy inclusion is proposed and its related properties are presented. The properties of fuzzy belief measures and fuzzy plausibility measures determined by the lower and upper approximation of fuzzy rough set are discussed and applied in attribute reduction of random fuzzy information system. 2010 Vol. 23 (4): 531-538 [Abstract] ( 285 ) [HTML 1KB] [ PDF 467KB] ( 561 ) 539 Statistical Gait Recognition Based on Tangent Angle Features ZHANG Yuan-Yuan,WU Xiao-Juan,RUAN Qiu-Qi A statistical gait recognition algorithm based on tangent angle features is proposed in this paper. Firstly, the method of Procrustes shape analysis is used to produce Procrustes compact Mean Shape (PMS) from the continuous posture variation of human body profile outline in gait sequence. The PMS is utilized as the primitive gait feature in this paper. Then the tangent angle corresponding to the tangential vector at each sample point on the PMS is computed. The tangent angle is considered to reflect the local appearance and tendency at that particular point of the outline and is treated as a local discriminative gait feature called Tangent Angle Feature (TAF). Finally, the Local Tangent angle Dissimilarity is used to measure the distance between two different TAFs, and the simplest standard classifiers are used to implement gait recognition. The experimental results on CASIA database and SOTON database show that the proposed algorithm is simple and effective and outperforms other existing approaches in terms of recognition accuracy. 2010 Vol. 23 (4): 539-545 [Abstract] ( 305 ) [HTML 1KB] [ PDF 499KB] ( 793 ) 546 k-Nearest-Neighbor Network Based Data Clustering Algorithm JIN Di,LIU Jie,JIA Zheng-Xue,LIU Da-You Data clustering is a hotspot in data mining area. Though there have been lots of data clustering algorithms now, the clustering accuracy of them is far from perfect. A structural similarity based network clustering algorithm (SSNCA) is proposed in this paper, which attempt to further improve the data clustering accuracy from the view of network clustering. The concrete solution scheme is that vector dataset for clustering is converted to a k-Nearest-Neigborhood network and SSNCA is used to cluster this network. Comparing SSNCA with the algorithms of c-Means and affinity propagation (AP), experimental result shows that the fitness value got by the proposed algorithm is a little worse than AP, but its clustering accuracy is obviously better than that of the other two algorithms. 2010 Vol. 23 (4): 546-551 [Abstract] ( 362 ) [HTML 1KB] [ PDF 422KB] ( 936 ) 552 An Automatic Ear Detection Method Based on Improved GVF Snake LI Yi-Bo,HUANG Zeng-Xi,ZHANG Hai-Jun,MU Zhi-Chun In recent years the biological characteristic research of human ear can only rely on manual positioning and segmentation, which is one obstruction in the practice process of human ear recognition technology. A method of human ear automatic detection is presented in this paper. Firstly, it detects the human ear pieces by taking advantage of YCbCr Skin-Color Model and Gentle AdaBoost cascade classifier, and then uses the improved GVF Snake method to extract the external ear contour. The method extracts the initial contour very close to the actual edge of human ear by building up the ear-shaped map. Accordingly, it not only saves iteration time, but also improves the accuracy of external ear detection in way of GVF Snake. About 97.3% of the correct detection rate is obtained on USTB ear database in our experiment. The experiment results show that the approach has better detection performance and robustness. 2010 Vol. 23 (4): 552-559 [Abstract] ( 372 ) [HTML 1KB] [ PDF 642KB] ( 649 ) 560 Improved Artificial Fish Swarm Algorithm for Vehicle Routing Problem with Backhaul and Fuzzy Demand LIU Yi The vehicle routing problem with backhaul and fuzzy demand (VRPBFD) is one of the most important and difficult problems in operational research filed. A mathematical model for the problem is built in this paper and the improved artificial fish swarm algorithm is proposed. With the effective combination of bionic principle in artificial fish swarm algorithm and subjective preference from decision maker, the optimization function is reconstructed. The optimization ability is raised by dynamically adjusting moving steps of artificial fishes, visual range and neighborhood values. The experiment results of simulation show that the improved algorithm has the validity and superiority. 2010 Vol. 23 (4): 560-564 [Abstract] ( 366 ) [HTML 1KB] [ PDF 337KB] ( 529 ) 565 Pairwise Diversity Measures Based Selective Ensemble Method YANG Chang-Sheng,TAO Liang,CAO Zhen-Tian,WANG Shi-Yi Effective generating individual learners with strong generalization ability and great diversity is the key issue of ensemble learning. To improve diversity and accuracy of learners, Pairwise Diversity Measures based Selective Ensemble (PDMSEN) is proposed in this paper. Furthermore, an improved method is studied to advance the speed of the algorithm and support parallel computing. Finally, through applying BP neural networks as base learners, the experiment is carried out on selected UCI database and the improved algorithm is compared with Bagging and GASEN (Genetic Algorithm based Selected Ensemble) algorithms. Experimental results demonstrate that the learning speed of the proposed algorithm is superior to that of the GASEN algorithm with the same learning performance. 2010 Vol. 23 (4): 565-571 [Abstract] ( 384 ) [HTML 1KB] [ PDF 506KB] ( 680 ) 572 Minimum Generation Error Training Based on Perceptually Weighted Line Spectral Pair Distance for Statistical Parametric Speech Synthesis LEI Ming,LING Zhen-Hua,DAI Li-Rong A Minimum Generation Error (MGE) training method based on perceptually weighted Line Spectral Pair (LSP) distance is proposed to improve the performance of Hidden Markov Model (HMM) based parametric speech synthesis system. The generation error defined by Euclidean distance used in the traditional MGE training, is not eligible in measuring the real gap between generated spectrum and natural spectrum when the speech spectrum is described by LSP. Although using generation error defined by Log Spectral Distortion (LSD) having nothing to do with spectrum parameters manages to deal with this problem, the improvement seems trivial compared to the incurred higher computational complexity. In this paper, an MGE training criterion based on weighted LSP distance is proposed, and this MGE training method is subjectively and objectively contrasted with different weighted methods and LSD based MGE training method. Eventually, a perceptually weighted training method is obtained, which not only achieves the best performance, but also incurs no extra computational complexity compared with the traditional MGE training. 2010 Vol. 23 (4): 572-579 [Abstract] ( 361 ) [HTML 1KB] [ PDF 608KB] ( 713 ) 580 Mean Shift Dynamic Deforming Hand Gesture Tracking Algorithm Based on Region Growth ZHANG Qiu-Yu,HU Jian-Qiang,ZHANG Mo-Yi Considering the traditional Mean shift algorithm has the problem of low tracking accuracy owing to the background pixels changing in searching window and that of high time complexity led by the updating of hand gesture model, a dynamic deforming hand gesture tracking algorithm based on the integration of region growth and Mean shift is proposed in this paper. The initiation of hand gesture center is automatically accomplished by the frames difference method at the initial tracking stage. Then, the hand gesture pixel samples are gathered by the region growth method. Finally, the position of the object center is accurately located by the Mean shift. Experimental results show that the proposed method can track the dynamic deforming hand gestures accurately in real time, reduce the time complexity of algorithm and make certain the stability and the continuity of the dynamic object tracking. 2010 Vol. 23 (4): 580-585 [Abstract] ( 335 ) [HTML 1KB] [ PDF 420KB] ( 581 ) 586 Particle Swarm Optimization Algorithm Using Dynamic Neighborhood Adjustment CHEN Zi-Yu,HE Zhong-Shi,ZHANG Cheng To keep a balance between global searching ability and searching speed, a particle swarm optimization algorithm using dynamic neighborhood adjustment (PSODNA) is presented. According to swarm diversity variation and evolutionary state, neighborhood structure of the particle swarm is dynamically changed in PSODNA. Population entropy is introduced to estimate swarm diversity. Particle neighborhood extension factor and local effect factor are defined to describe the evolutionary state. And neighborhood expansion and constraint strategies are proposed to control the influence of good particles. The experimental results show that the proposed algorithm has great superiority both in global searching ability and searching speed. 2010 Vol. 23 (4): 586-590 [Abstract] ( 372 ) [HTML 1KB] [ PDF 447KB] ( 630 )

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