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....

2009 Vol.22 Issue.5, Published 2009-10-25 Papers and Reports    Researches and Applications    Surveys and Reviews        Papers and Reports 673 Speech Emotion Recognition Based on Covariance Descriptor and Riemannian Manifold LIU Jia, CHEN Chun, YE Cheng-Xi, LI Na, BU Jia-Jun An algorithm for speech emotion recognition is proposed based on covariance descriptor and Riemannian manifold. According to the extracted acoustic features, covariance matrices are computed as the emotion descriptors of sentences. With the consideration of high dimensional characteristic of the space constructed by non-singular covariance matrices, an affine invariance metric is adopted to make the space meet the requirement of Riemannian manifold. With differential geometry, the speech emotion recognition is performed on the manifold. The experimental results show a significant improvement in recognition accuracy, especially under noisy environments. 2009 Vol. 22 (5): 673-677 [Abstract] ( 414 ) [HTML 1KB] [ PDF 443KB] ( 703 ) 678 Analysis of Structural Clustering Based on Normalized Metric TANG Xu-Qing, ZHU Ping, CHENG Jia-Xing On the basis of the ordered granular space, structural clustering (or classification) analysis is proposed based on normalized metric. Firstly, the concept of consistent clustering according to metric is presented, and the research on consistent clustering characteristic of ordered granular space is given. Secondly, structural clustering analysis theory is given based on normalized metric, and the algorithm to obtain its structural clustering is discussed. Thirdly, the research on the determination of the optimal clustering based on ordered granular space is carried out. A method to obtain the optimal clustering is given, and the method is global optimal. Finally, the fusion technology based on structural clusters of normalized metrics is studied by the intersection operation of two normalized metrics. The conclusions provide a comprehensive theory and methodology on structural clustering (or classification) analysis based on metric. 2009 Vol. 22 (5): 678-688 [Abstract] ( 313 ) [HTML 1KB] [ PDF 471KB] ( 400 ) 689 Region-Level Moving Object Detection Method WANG Huan, REN Ming-Wu, YANG Jing-Yu Conventional moving object detection methods are usually based on pixel or hard-divided-blocks. In this paper, watershed transformation is applied to segment an image into homogenous regions adaptively, and these regions are then used as minimal processing cell for moving object detection. To alleviate the over-segmentation problem, watershed transformation is performed on multi-stage morphological gradient image. To meet the requirement of low false alarm and high real-time performance, a homogenous region belonging to foreground or background is judged directly by measuring intensity difference, chromaticity distortion, and intensity relation among adjacent regions between the region and its corresponding regions in a series of maintained history background image. The false alarm of the proposed approach is lower than that of the popular detection methods, and it avoids the hard block-split problem in region-based detection approaches. The processing speed of the proposed approach is also satisfactory. Experiments on several benchmark video sequences are made including both indoor and outdoor scenes, and the results demonstrate the effectiveness of the proposed algorithm. 2009 Vol. 22 (5): 689-696 [Abstract] ( 324 ) [HTML 1KB] [ PDF 1044KB] ( 437 ) 697 Rough Set Model Based on Logical OR Operation of Precision and Grade ZHANG Xian-Yong, XIONG Fang, MO Zhi-Wen Precision and grade are two important indexes for quantitative research. The purpose of this paper is to combine precision and grade and explore a new extended rough set model. Transformation formulas of variable precision approximations and graded approximations are obtained by studying the relationship between them. Based on the logical OR requirement of precision and grade, rough set model of logical OR operation of precision and grade and new rough set regions are proposed. In rough set model of logical OR operation of precision and grade, basic structures of rough set regions are gained. The regular algorithm and structure algorithm are proposed and analyzed to calculate rough set regions. The variable precision rough set model, graded rough set model and classical rough set model are extended by the proposed model, and thus these models get corresponding structures of rough set regions. 2009 Vol. 22 (5): 697-703 [Abstract] ( 261 ) [HTML 1KB] [ PDF 315KB] ( 569 ) 704 An Adaptive Pedestrian Tracking Algorithm with Prior Knowledge CHENG You-Long, LI Bin, ZHANG Wen-Cong, ZHUANG Zhen-Quan In actual surveillance conditions, many uncertainties exist in pedestrian movement. These movements may disturb the current tracking algorithms and result in tracking lost. An adaptive pedestrian tracking algorithm is proposed. In this algorithm, the prior knowledge of pedestrian detection is embedded into the self-learning process of object model. Firstly, offline training is performed to get a set of sub-classifiers with strong discriminability and prior knowledge of the pedestrians. Then, online boosting algorithm is used for learning and updating the pedestrian's dynamic model from the offline trained sub classifier set. Experimental results show that the proposed method efficiently relieves the conflict between adaptation and drifting, and tracks pedestrian with various uncertain movement under the actual surveillance conditions. 2009 Vol. 22 (5): 704-708 [Abstract] ( 312 ) [HTML 1KB] [ PDF 1458KB] ( 679 ) 709 Documents Sampling Based on Feature Selection and Condensing Techniques HAO Xiu-Lan, TAO Xiao-Peng, WANG Shu-Yun, XU He-Xiang, HU Yun-Fa As an instance based classifier, kNN has many computational and store requirements. Meanwhile, the poor performance of kNN classifier is caused by the imbalance distribution of training data. Aiming at these defects of kNN classifier, a technique, combining feature selection and condensing, is proposed to reduce the time cost and the space of kNN classifier. The proposed algorithm is divided into two steps. Firstly, several traditional methods of feature selection are combined to form features for each class. Then, redundant cases are removed by combination of class features contained in samples with Condensing algorithm. Experimental results indicate when the sample set acquired by the proposed method is used as training set, the classifier saves the time cost and the space dramatically, and the performance of the kNN classifier is improved because noisy data are removed from the training set. 2009 Vol. 22 (5): 709-717 [Abstract] ( 344 ) [HTML 1KB] [ PDF 612KB] ( 791 ) 718 An AMPF and FastSLAM Based Compositive SLAM Algorithm ZHOU Wu, ZHAO Chun-Xia, ZHANG Hao-Feng An AMPF and FastSLAM Based Compositive SLAM algorithm is presented to improve the performance of samples and increase the estimation accuracy. The auxiliary marginal particle filter (AMPF) is combined with the FastSLAM framework. In the proposed algorithm, the robot pose is estimated by AMPF, and the proposal pose distribution of each particle is estimated by UKF. Appropriate sampling strategy and particle data structure are designed to be compatible with both AMPF and FastSLAM framework. The map is estimated recursively by EKF in the FastSLAM framework. Experimental results indicate that the samples' performance of the proposed algorithm is better than that of FastSLAM 2.0, and the estimation accuracy of the proposed algorithm is higher than that of FastSLAM 2.0. Moreover, the estimation accuracy of the proposed algorithm is good with few samples. Therefore, it is feasible to improve the computational efficiency by reducing the number of samples. 2009 Vol. 22 (5): 718-725 [Abstract] ( 344 ) [HTML 1KB] [ PDF 667KB] ( 968 ) 726 An Improved PSO Method Based on Adaptive Cognitive Domain LIU Tun-Dong, CHEN De-Bao, LI Su-Wen, WANG Ying To improve the convergent performance of particle swarm optimization(PSO), an adaptive cognitive domain particle swarm optimization(ACDPSO) method is proposed. In the updating equations of particles, the current best position, which the particle achieves, is determined by the center of the best calculated position and the cognizant direction of the particle. Linear decreasing inertia weight is used to optimize particles. Three different PSOs, particle swarm with constant weight(CWPSO), linear decreasing inertia weight PSO(LDWPSO) and Ladder PSO(LPSO), are combined with the proposed method to test the performance of the proposed method, and the results indicate that the proposed method is effective. 2009 Vol. 22 (5): 726-730 [Abstract] ( 339 ) [HTML 1KB] [ PDF 513KB] ( 464 ) 731 EAPSC: Efficient Clustering of Skyline Objects HUANG Zhen-Hua, XIANG Yang, LIN Chen A concept, SkyCluster, is proposed. It clusters the skyline objects according to their associative distance. The skyline query and cluster processing are all CPUsensitive. Hence, to improve the efficiency of obtaining SkyClusters, an efficient approach, EAPSC, is presented to cluster skyline objects. EAPSC algorithm is based on the novel index tree SLT and employs several interesting properties of SLT to produce SkyClusters fast. Furthermore, the theoretical analysis and experimental results demonstrate the proposed method is efficient and effective. 2009 Vol. 22 (5): 731-734 [Abstract] ( 225 ) [HTML 1KB] [ PDF 311KB] ( 537 ) Surveys and Reviews 735 Some Developments on Semi-Supervised Clustering LI Kun-Lun, CAO Zheng, CAO Li-Ping, ZHANG Chao, LIU Ming Small amount of labeled data are used in semi-supervised clustering algorithms to improve the performance of the algorithms. It is a research hotspot in pattern recognition and its related fields. In this paper, some developments on semi-supervised clustering are introduced including constraint-based, distance-based and the combination of them. Using semi-supervised strategy to fuzzy C-means, a semi-supervised fuzzy C-means (constrained FCM) algorithm is proposed. Experimental results show that the proposed method obtains better accuracy compared with FCM and semi-supervised K-means. 2009 Vol. 22 (5): 735-742 [Abstract] ( 388 ) [HTML 1KB] [ PDF 413KB] ( 1201 ) Researches and Applications 743 Low-Level Image Features Based Human Body Detection Using Hidden Markov Model XU Cui, ZHENG Ying, WANG Zeng-Fu A method for human body detection from single image is presented. A hidden Markov model (HMM) is used to represent the human body. Based on the given series of human body configuration, the best image segments are inferred. Thus, the problem of human body detection is transformed into a HMM decoding one. Firstly, the image is segmented using Mean-Shift based procedure and the torso regions are searched according to color information. Secondly, the low-level features of shading, color and contour are combined to estimate the probability of feature matching and find the limb candidates. Finally, the connection probabilities of candidates are computed and the best fit human body regions are inferred by HMM decoding algorithm. The experimental results indicate that the proposed detection method detects reasonable human body well even from images with complex background and various pose. Compared with other detection methods, the proposed method approximates the body parts by rectangles and gets the integrally segmented human region. Moreover, it adapts to the low resolution images or images with people who are small or suffer from motion blur. 2009 Vol. 22 (5): 743-749 [Abstract] ( 304 ) [HTML 1KB] [ PDF 1421KB] ( 965 ) 750 Co-Training Semi-Supervised Active Learning Algorithm with Noise Filter ZHAN Yong-Zhao, CHEN Ya-Bi The classification performance of the classifier based on semi-supervised learning is weakened when the noise samples are introduced. An algorithm called co-training semi-supervised active learning with noise filter is presented to overcome this disadvantage. In this algorithm, three fuzzy buried Markov models are used to perform semi-supervised learning cooperatively. Some human-computer interactions are actively introduced into labelling the unlabeled sample at certain time in order to avoid the rejective judgment when the classifiers do not agree with each other and the inaccurate judgment when the initial weak classifiers all agree. Meanwhile, the noise filter is used to filter the possible noise samples which are labeled automatically by the computer. The proposed algorithm is applied to facial expression recognition. The experimental results show that the algorithm can effectively improve the utilization of unlabeled samples, reduce the introduction of noise samples and raise the accuracy of expression recognition. 2009 Vol. 22 (5): 750-755 [Abstract] ( 311 ) [HTML 1KB] [ PDF 448KB] ( 768 ) 756 Methods for the Extension Rules of Ontology Based on Multidimensional Association Rules DONG Jun, WANG Suo-Ping, XIONG Fan-Lun, ZHANG You-Hua Currently, the extension and enrichment for ontology have some limitations. Therefore, an approach is presented to extend ontology rules with multi-dimensional association rule technology. The conception ontology is enriched and extended by ontology rules extraction, consistency treatment under guidance of the ontology, rules mapping establishment, and the re-identification and update for conception ontology. The experimental results of tea diseases and pests predicting ontology show that the proposed approach can be easily implemented and has good feasibility and validity. 2009 Vol. 22 (5): 756-762 [Abstract] ( 272 ) [HTML 1KB] [ PDF 383KB] ( 482 ) 763 An Image Algebraic Reconstruction Technique Based on POCS Restriction HU Xiao-Zhou, KONG Bin, CHENG Er-Kang, HU Rong-Xiang Algebraic reconstruction based on incomplete projection data is a hot issue in CT application. An improved algebraic reconstruction technique (ART) is proposed based on the analysis of relationships between images with mutual perpendicular projection angles. The projection coefficient matrix is calculated by recording the indices of ray-cross grids and the lengths of grid-ray intersections. In the process of reverse projection, POCS restriction is used to reconstruct the image from the incomplete data. The experimental results show that compared with the ART algorithm, the proposed algorithm greatly improves the speed and the quality of image reconstruction. 2009 Vol. 22 (5): 763-768 [Abstract] ( 303 ) [HTML 1KB] [ PDF 647KB] ( 483 ) 769 Kernel Based Class-Wise Non-Locality Preserving Projection WANG Wen-Jun, ZHANG Jun-Ying A feature extraction method is proposed, namely class-wise non-locality preserving projection (CNLPP). The kernelized counterpart of CNLPP linear feature extractor is also established. Based on the linear feature extractor-non-locality preserving projection (NLPP), CNLPP utilizes between-class information to guide the procedure of feature extraction. CNLPP takes both the relation information and the class information into account. A kernel version of CNLPP, namely Kernel based CNLPP (KCNLPP), is developed by applying the kernel trick to CNLPP to enhance its performance on nonlinear feature extraction. Experiments on yeast gene expression data and NCI gene expression data are performed to test and evaluate the performance of the proposed algorithm, and the results show that KCNLPP achieves relatively high recognition accuracy. 2009 Vol. 22 (5): 769-773 [Abstract] ( 268 ) [HTML 1KB] [ PDF 295KB] ( 456 ) 774 Image Retrieval Based on Transductive Support Vector Machine CHEN Shi, GUO Mao-Zu, LIU Yang, DENG Chao To reduce the gap between low-level image features and high-level semantic concept, support vector machine based relevance feedback draws more and more attentions. However, the information embedded in unlabeled samples is not utilized in that method. In order to exploit these information sufficiently, the transductive support vector machine (TSVM) is introduced into feedback process. Based on analyzing the characters of feature vector for TSVM, a color sparse feature is designed as the image description feature combined with the texture feature. Experimental results show that the proposed method is more discriminative than the feedback process using support vector machine (SVM), and TSVM obtains good results when applied to other fields. 2009 Vol. 22 (5): 774-779 [Abstract] ( 268 ) [HTML 1KB] [ PDF 372KB] ( 582 ) 780 Document Cluster Ensemble Algorithms Based on Matrix Spectral Analysis XU Sen, LU Zhi-Mao, GU Guo-Chang Cluster ensemble techniques are effective in improving both the robustness and the stability of the single clustering algorithm. How to combine multiple clusters to yield a final superior clustering result is critical in cluster ensemble. Spectral clustering algorithm is introduced to solve document cluster ensemble problem. Normalized Laplacian matrix-based spectral algorithm (NLMSA) is proposed. According to algebraic transformation, it computes eigenvalues and eigenvectors of a small matrix to obtain the eigenvectors of normalized Laplacian matrix. The key idea of spectral clustering algorithm is further investigated, and hyperedge transition matrix-based spectral algorithm (HTMSA) is proposed. It attains the low dimensional embeddings of documents by those of hyperedges and then the K-means algorithm is used to cluster according to those embedding results of documents. Experimental results on TREC and Reuters document sets demonstrate the effectiveness of the proposed algorithms. Both NLMSA and HTMSA outperform other cluster ensemble techniques based on graph partitioning. NLMSA obtains better results than HTMSA while the computational cost of HTMSA is much lower than that of NLMSA. 2009 Vol. 22 (5): 780-786 [Abstract] ( 301 ) [HTML 1KB] [ PDF 410KB] ( 895 ) 787 Fast Multi-Stage Hybrid Fingerprint Matching CAO Guo, MAO Zhi-Hong, MEI Yuan, SUN Quan-Sen A hybrid multi-stage fingerprint matching method is proposed. After the extraction of the image features and the construction of the fingerprint minutiae pattern, a multi-stage fingerprint matching procedure is performed. Firstly, the Euclidean distance between the two corresponding image features is calculated and the preliminary matching is realized by comparing the Euclidean distance. Then, the query minutiae pattern with the template minutiae pattern are directly matched at the second matching stage. Finally, further matching is carried out using multiple pairs of reference minutiae obtained at the second stage. Experimental results indicate that the proposed multi-stage matching method is fast and effective. 2009 Vol. 22 (5): 787-793 [Abstract] ( 325 ) [HTML 1KB] [ PDF 1195KB] ( 452 ) 794 Variable Universe Fuzzy Control Based on Multilayer Ant Colony Algorithm ZHAO Yun-Tao, WANG Jing, XIE Xin-Liang, HOU Qiang Aiming at the low accuracy and poor adaptation of the fuzzy control algorithm, an adaptive fuzzy controller with variable universe is proposed. Firstly, based on the theory of ant colony optimization (ACO), a multilayer ant colony algorithm for continuous domains is designed. Through decomposing solution space into finite grids, the proposed algorithm is realized by three-stage search strategy with the progress of iterations, and heterogeneous mechanism is utilized at each stage. Secondly, the intelligent algorithm is used to optimize the contraction and expansion factors. Thus, the universe of fuzzy controller is adjusted on line according to performance indicator. Meanwhile, the adaptive fuzzy controller with variable universe is adopted to hydraulic servo system of medium plate. Finally, the simulation results show that the convergence speed of the system with adaptive fuzzy controller is higher than that of others. And the proposed control strategy has fine property in effectiveness and robustness. 2009 Vol. 22 (5): 794-798 [Abstract] ( 388 ) [HTML 1KB] [ PDF 491KB] ( 862 ) 799 Boundary Points Detection Algorithm Based on Coefficient of Variation XUE Li-Xiang, QIU Bao-Zhi In order to detect boundary points of clusters effectively, an algorithm is proposed, namely boundary points detecting algorithm based on coefficient of variation(BAND). BAND computes the average distance between one object and its k-distance neighbors. The density of each object is obtained by the reciprocal of average distance. Then the boundary points are found by using the coefficient of variation to portray the distribution of data objects. The experimental results show BAND effectively detects boundary points on noisy datasets with clusters of arbitrary shapes, sizes and different densities. 2009 Vol. 22 (5): 799-802 [Abstract] ( 364 ) [HTML 1KB] [ PDF 581KB] ( 713 ) 803 Clustering Method of Time Series Based on EMD and K-means Algorithm LIU Hui-Ting, NI Zhi-Wei Dimension reduction of time series and noise in sequences filtering are important prerequisites for effective realization of time series clustering. A method is proposed to preprocess time series effectively. Firstly, the trend of a time sequence is got by using empirical mode decomposition method. Then, the trend series are divided into several segments by bottom-up algorithm. Finally, the piecewise series are translated into uniform sequences, and each of them is composed of -1, 0 and 1. To prove that the proposed method can achieve dimensionality reduction and filter out the noise from the data sequence, K-means algorithm is utilized to finish clustering of pretreated time series. Experimental results show clustering of pretreated data sequences is better than that of the original series. 2009 Vol. 22 (5): 803-808 [Abstract] ( 528 ) [HTML 1KB] [ PDF 819KB] ( 677 )

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