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....

2006 Vol.19 Issue.5, Published 2006-10-25 Papers and Reports    Researches and Applications        Papers and Reports 561 Learning Bayesian Networks Structure with Hidden Variables WANG ShuangCheng, LIU XiHua, TANG HaiYan At present the method of learning Bayesian network structure with hidden variables is mainly based on the scoringsearch method combined with EM algorithm. But it is inefficient and unreliable. A new method of learning Bayesian network structure with hidden variables is presented. In this method, the Bayesian network structure without hidden variables is set up based on basic dependency relationship between variables and basic structures between nodes and dependency analysis idea. Hidden variables are found in terms of the dimension of cliques in the moral graph of Bayesian network. The value, the dimension and the local structure of hidden variables are made based on dependency strcture between variables, Gibbs sampling and MDL criterion. The method can avoide the exponential complexity of standard Gibbs sampling and the main problems of the existing algorithm of learning Bayesian network structure with hidden variables. Experimental results show that this algorithm can effectively learn Bayesian network strcture with hidden variables. 2006 Vol. 19 (5): 561-566 [Abstract] ( 305 ) [HTML 1KB] [ PDF 411KB] ( 430 ) 567 Language Identification between Mandarin and English with State Duration Information SUN Jian, WANG ZuoYing Different languages have different pronunciation rates, so the state duration reflects the pronunciation rate of a language. The phone recognition system and LVCSR (Large Vocabulary Continuous Speech Recognition) system are developed by using DDBHMM (Duration Distribution Based Hidden Markov Model). Both systems are used to identify Mandarin and English. The results prove that DDBHMM describes the state duration accurately and improves the performance of language identification. 2006 Vol. 19 (5): 567-571 [Abstract] ( 280 ) [HTML 1KB] [ PDF 369KB] ( 455 ) 572 An Efficient Attribute Reduction Algorithm Based on Discernibility Object Pair Set XU ZhangYan, YANG BingRu, SONG Wei The definition of discernibility object pair set and the corresponding definition of attribute reduction are introduced. It is proved that the definition of attribute reduction is equivalent to the one based on positive region. Since U/C is important for computing the discernibility object pair set, an algorithm for computing U/C is designed, whose time complexity is cut down to O(｜C｜｜U｜). Under this condition, an efficient attribute reduction algorithm is proposed, whose time and space complexity are cut down to O(｜C｜｜U｜)+O(｜C｜(｜U/C｜2)) and O(｜U｜)+O(｜U/C｜2) respectively. Finally, an example is used to illustrate the efficiency of the new algorithm. 2006 Vol. 19 (5): 572-577 [Abstract] ( 218 ) [HTML 1KB] [ PDF 308KB] ( 400 ) 578 Confusable Chinese Speech Recognition Based on HMM/SVM TwoLevel Architecture WANG HuanLiang, HAN JiQing, LI HaiFeng, ZHENG TieRan The recognition rate for confusable speech is still low in stateoftheart Chinese speech recognition systems based on HMM. The inherent defects of HMM are analyzed, then a twolevelarchitecture recognition framework combining HMM and SVM is proposed. A confidence estimation module is adopted to improve the performance and efficiency of the system. The information obtained by Viterbi decoding is utilized to construct new classes of feature for SVM, which solves the problem that the conventional SVM cannot directly process variable length sequences. The relevant issues, such as confidence estimation, classification feature extraction and SVM recognizer construction, are addressed. The experimental results of confusable Chinese speech show that compared with the hybrid HMM/SVM based system the proposed method can highly improve the recognition rate with little impact on the running speed. 2006 Vol. 19 (5): 578-584 [Abstract] ( 351 ) [HTML 1KB] [ PDF 633KB] ( 535 ) 585 Block Statistics Based Gabor Feature Representation and Its Application to Face Recognition LONG Fei, YE XueYi, LI Bin, YAO Peng, ZHUANG ZhenQuan Face representation based on Gabor features has attracted much attention and achieved great success in face recognition for some favorable attributes of Gabor wavelets such as spatial locality and orientation selectivity. A large number of Gabor features are produced with varying parameters in the position, scale and orientation of filters. In some existing methods, useful discriminatory information may be lost due to downsampling Gabor features directly. To reduce the loss, a block statistics based Gabor feature representation method is proposed. The effectiveness of this method is demonstrated by template matching test on ORL face database, and the comparative test results show that this method can yield better recognition accuracy with much fewer Gabor features as well as less CPU time of feature matching than the existing approach of downsampling based Gabor feature representation. In addition, Generalized Discriminant Analysis (GDA) which performs dimensionality reduction to Gabor features is used to produce more compact and discriminatory face representation. The experimental results of face recognition using different similarity measures show that the proposed method outperforms the famous Eigenfaces and Fisherfaces methods significantly, and the rationality of this combination is also comparatively demonstrated. 2006 Vol. 19 (5): 585-590 [Abstract] ( 257 ) [HTML 1KB] [ PDF 788KB] ( 588 ) 591 Mapping Algorithm of Mobile Robot in Unknown Indoor Environment Based on Active Exploration ZHOU GuangMing, CHEN ZongHai, LIU NianQing, JIA MengLei The mobile robot mapping in unknown indoor environment based on sonar sensors is discussed. Aiming at the inherent limitations such as poor azimuth resolution and specular reflection, an accurate local environment model is achieved by analyzing the relationship of adjacent sonar information, and an active exploration strategy for robot mapping is presented based on this model. Comparative experiments are conducted on a real mobile robot in indoor environment using the proposed mapping algorithm and another popular mapping algorithm. Experimental results show that the proposed algorithm can overcome the inherent limitation of sonar sensor and get a complete map of the indoor environment accurately and effectively. 2006 Vol. 19 (5): 591-597 [Abstract] ( 289 ) [HTML 1KB] [ PDF 974KB] ( 411 ) 598 Least Squares Support Vector Machine Based on Scaling Kernel Function WU FangFang, ZHAO YinLiang The kernel function of support vector machine (SVM) is an important factor to the learning result of SVM. Based on the wavelet decomposition and conditions of the support vector kernel function, a new scaling kernel function for SVM (SSVM) is proposed. This function is not only a kind of horizontal floating orthonormal function, but also can simulate any curve in quadratic continuous integral space, thus it enhances the generalization ability of the SVM. According to the scaling kernel function and the regularization theory, a least squares support vector machine on scaling kernel function (LSSSVM) is proposed to simplify the solving process of SSVM. The LSSSVM is then applied to the regression analysis and classification. Experimental results show that the precision of regression is improved, compared with LSSVM whose kernel function is Gauss function. 2006 Vol. 19 (5): 598-603 [Abstract] ( 308 ) [HTML 1KB] [ PDF 664KB] ( 465 ) 604 Bayesian Network Structure Refinement Method JIANG GuoPing, CHEN YingWu Aiming at the twostage modeling process, a Bayesian network structure refinement method based on conditional mutual information measure is put forward. The detailed procedure of this method is presented, the correctness is proved and the computing complexity is analyzed. A wellknown Bayesian networks-Alarm is applied to show its validity. 2006 Vol. 19 (5): 604-610 [Abstract] ( 389 ) [HTML 1KB] [ PDF 413KB] ( 473 ) 611 A Discretization Method Based on Artificial FishSwarm Algorithm ZENG JianWu, ZHANG JianMing, WANG ShuQing A new method of optimal discretization is proposed in order to solve the illconditioned problem caused by inefficient discretization approaches. The artificial fishswarm algorithm is used to optimize the objective function converted by the problem of dividing the segmental points. The virtual segmental points are introduced to effectively combine the discrete intervals based on the swarm deed of fishswarm algorithm. The case analysis shows that the number of the discrete intervals is small. Furthermore, the decision rules derived from the proposed method are simple and with fine robustness. 2006 Vol. 19 (5): 611-616 [Abstract] ( 277 ) [HTML 1KB] [ PDF 351KB] ( 427 ) 617 The Automated Geometry Reasoning Network Based on Equivalent Class Reasoning JIANG JianGuo, ZHANG JingZhong To improve the reasoning efficiency of the inference engine, a new inference engine, namely automated geometry reasoning network is presented, into which the Rete pattern matching algorithm and the equivalent class reasoning technique are integrated. The new inference engine is implemented with Lisp and tested with more than 50 nontrivial geometry theorems. The experimental results show that it is more efficient. 2006 Vol. 19 (5): 617-628 [Abstract] ( 393 ) [HTML 1KB] [ PDF 331KB] ( 459 ) Researches and Applications 629 PCA in Speech Detection ZHU JunBo, ZHU XiaJun, WANG ShouJue It is essential for speech processing system to have robust speech detection. In this paper, a PCA(principal component analysis)based speech detection method is proposed. A good result of the examination by using this method is gotten. In this method, speech and nonspeech subspaces are created respectively by using PCA. The result of fast PCA is the basis of the new subspace. By analysis the distribution of the data in subspace, the speech and nonspeech can be detected respectively. Creating a number of different type nonspeech subspaces can get a better performance than creating one. 2006 Vol. 19 (5): 629-633 [Abstract] ( 343 ) [HTML 1KB] [ PDF 406KB] ( 438 ) 634 Automatic News Audio Classification Method Based on Selective Ensemble SVMs HAN Bing, GAO XinBo, JI HongBing As a significant clue for video indexing and retrieval, audio detection and classification has attracted much attention and become a hot topic. On the basis of the prior model of news video structure, a selective ensemble support vector machines (SENSVM) is proposed to detect and classify the news audio into 4 types: silence, music, speech, and speech with music background. Experiments on real news audio clips of 8514s in total length illustrate that the average accuracy rate of the proposed audio classification method reaches 98.2%, which is much better than that of the available SVMbased method or the traditional thresholdbased method. 2006 Vol. 19 (5): 634-639 [Abstract] ( 292 ) [HTML 1KB] [ PDF 365KB] ( 445 ) 640 EdgeBased Moving Shadow Removal Algorithm for Indoor Video Sequence XIAO Mei, HAN ChongZhao Because the moving shadows usually were extracted along with objects by motion detection algorithms, an edgebased approach to remove moving shadows for indoor sequences is proposed. Firstly, edges of input images are found using Canny operator and the input image is segmented according to gradient magnitude. Next, the moving edges of real moving object are obtained using moving edge properties, then the part boundaries of real foregrounds near the moving edges are found. Finally, the complete foreground objects are constructed by means of the proposed border tracing technique, thus the moving object detection exactitude is improved. Simulation results show that the proposed method can effectively separate the moving objects from their shadows caused by different distances of the lamphouse, diverse shadows directions and foreground colors. 2006 Vol. 19 (5): 640-644 [Abstract] ( 261 ) [HTML 1KB] [ PDF 523KB] ( 486 ) 645 Feature Extraction and Image Reconstruction of Video Sequence Based on Nonlinear Dimensionality Reduction Algorithms LI Hua, DU SiDan, LU Fan, GAO DunTang A dualchannel extension of nonlinear dimensionality reduction algorithm is proposed according to the characteristic of stereo video sequence. In order to construct a mapping between high and low dimensional spaces, kneighbor kernel function method is presented, which solves the problem of lack of reconstruction algorithm and provides a new solution to video compression. Experimental results on several typical sequences show the advantages of the proposed approaches. 2006 Vol. 19 (5): 645-651 [Abstract] ( 314 ) [HTML 1KB] [ PDF 1173KB] ( 412 ) 652 Fast Algorithm for Face Contour Detection and Pose Estimation HU Yuan-Kui, WANG Zeng-Fu A method of face contour detection based on facial features and a fast method of face pose estimation are proposed. The contour detection method divides the face image into 9 regions according to the distribution of the facial features. In the selected regions, the crude face contours are detected and fitted by cubic polynomials. Finally, the complete face contour is achieved by connecting the contours in the selected regions. In order to estimate the facial 3D pose quickly and accurately, area model and approximate plane model are proposed according to the symmetry of the face. The experimental results indicate that the proposed contour detection method can achieve reasonable face contours even from face images with complex background and varying pose. Compared with other detection methods, the advantage of the proposed method is that the model is simpler and the computing is faster. 2006 Vol. 19 (5): 652-657 [Abstract] ( 428 ) [HTML 1KB] [ PDF 724KB] ( 792 ) 658 Segmentation Approach for Natural Images and Application to Object Detection in ViewBased Navigation HONG YiPing, YI JianQiang, ZHAO DongBin, LI XinZheng An improved mean shift segmentation approach for natural images is presented. And its application to object detection in viewbased navigation is proposed. In the improved segmentation, the color bandwidth derived from data analysis for clustering is estimated according to the image data. The density of each pixel is calculated based on the estimated bandwidth, and the local modes are searched by the direct density search way. Then a global standard for local mode merging is applied to get the final segmentation result. In its application to object detection, the object models constructed offline are used to verify the possible objects among the segmented regions. The experimental results indicate that the propesed approach can effectively detect the natural objects under complex backgrounds. 2006 Vol. 19 (5): 658-662 [Abstract] ( 284 ) [HTML 1KB] [ PDF 624KB] ( 442 ) 663 Texture Segmentation Based on Local Walsh Transform WANG XiaoMing A texture segmentation algorithm based on Local Walsh transform is presented. For each pixel in the texture images, the local Walsh transform coefficients are computed, and their 2nd, 4th and 6th order moments are estimated as the texture features of that pixel. Then all the pixels are clustered by fuzzy Cmeans algorithm. Experimental results of texture segmentation are fine. 2006 Vol. 19 (5): 663-666 [Abstract] ( 295 ) [HTML 1KB] [ PDF 1234KB] ( 356 ) 667 Improved Adaptive Algorithm of Blind Source Separation Based on Nonholonomic Natural Gradient NIU YiLong, WANG YingMin, WANG Yi Compared with the natural gradient learning algorithm of blind source separation, the nonholonomic natural gradient algorithm avoids numerical instability which is caused by the nonstationary source signals and the rapid magnitude change. Aiming at the difficulty in determining the nonlinear activation function, an improved algorithm using kurtosis to select activation function adaptively without available prior information is proposed. It retains the predominance of nonholonomic natural gradient algorithm in restoring nonstationary sources, and can be adapted to the sources of arbitrary distribution. Computer simulations show the performance of proposed method is better than that of the original algorithm with tangent function. 2006 Vol. 19 (5): 667-673 [Abstract] ( 308 ) [HTML 1KB] [ PDF 978KB] ( 507 ) 674 Graph Spectral Decomposition and Clustering KONG Min, CHEN SiBao, ZHAO HaiFeng, LUO Bin This paper explore how to use spectral methods for embedding and clustering unweighted graphs. The leading eignvectors of the graph adjacency matrix are employed to define eignmodes of the adjacency matrix. For each eigenmode, vectors of spectral properties are computed as feature vectors. These properties include the eigenmode perimeter, eigenmode volume, Cheeger number, intermode adjacency matrices and intermode edgedistance. Then these vectors are embedded in a patternspace using two contrasting approachs. The first of these involves performing principal or independent component analysis on the covariance matrix for the spectral pattern vectors. The second approach involves performing multidimensional scaling on the L2 norm for pairs of patten vectors. This paper also illustrate the utility of the embedding methods on neighbourhood graphs representing the arrangement of corner features in 2D images of 3D polyhedral objects. Experimental results show that clustering graphs using spectral properties of graphs is practical and effective. 2006 Vol. 19 (5): 674-679 [Abstract] ( 238 ) [HTML 1KB] [ PDF 876KB] ( 375 ) 680 Relation between Line Fitting Error and Direction of the Line in Binary Image LIU YangCheng, ZHU Feng The quantization error plays an important role in the field of image processing and computer vision. The relation between the error of line fitting and the line characters, especially the slope of the straight lines, is mainly analyzed. After a simple introduction of the line fitting, the definition of the sensitive space of straight lines is given, and its expression under different conditions is presented and proved. The emulational experiments suggest there is a close relation between the sensitive space and the error of fitting results. The error of intercept is carefully researched as the error of slope is much less than that of intercept. A method to evaluate intercept is presented. The experimental result shows the proposed method outperforms the traditional one. 2006 Vol. 19 (5): 680-684 [Abstract] ( 285 ) [HTML 1KB] [ PDF 525KB] ( 481 ) 685 Ear Recognition Method Based on Independent Component Analysis ZHANG HaiJun, MU ZhiChun, LIU Ke Ear recognition is a new biometric technique. Independent component analysis is applied to the feature extraction of ear images. And the nearest neighbor classifier, RBF neural networks classifier and SVM are combined with the ICA respectively for the classification. The experiments show that the ear recognition based on ICA has better performance compared with traditional method of principal component analysis. 2006 Vol. 19 (5): 685-688 [Abstract] ( 251 ) [HTML 1KB] [ PDF 335KB] ( 451 )

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