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.3, Published 2009-06-25 Papers and Reports    Researches and Applications        Papers and Reports 337 Improved Path Planning Based on Rapidly-Exploring Random Tree for Mobile Robot in Unknown Environment KANG Liang, ZHAO Chun-Xia, GUO Jian-Hui An improved path planning algorithm is proposed by combining rapidly-exploring random tree (RRT) and rolling path planning. In this algorithm, the real-time local environment information detected by the robot is fully used and the on-line planning is performed in a rolling style. Therefore, the RRT algorithm can be used in both known and unknown environment. Only the local environmental map is calculated in the planning to improve the planning efficiency, and thus the planning in real time is guaranteed. The calculation of analytical expressions of the obstacle can be ignored. Hence, the memory is saved greatly. Based on the algorithm of rapidly-exploring random, the heuristic evaluation function is introduced into the improved algorithm, so that the exploring random tree can grow in the direction of target point. The regression analysis, which avoids local minimum, enhances the capability of searching unknown space. The simulation results verify the effectiveness of the improved algorithm. 2009 Vol. 22 (3): 337-343 [Abstract] ( 444 ) [HTML 1KB] [ PDF 463KB] ( 2251 ) 344 Analysis of Second Order Motion SUN Bin, SANG Nong, LI Zheng, ZHENG Qing-Qing The study of the second order motion in biological vision is a new source of inspiration for algorithms and research directions in computer vision. In this paper, the second order motion can be divided into three typical group according to the modulation types: spatial modulate motion, temporal modulate motion and spatio-temporal modulate motion. Experiments are conducted on the first order motion perception based on correlation model and the second order motion perception by correlation model preceded with a nonlinear process called texture grabber. The computational results are consistent with the previous suggestion that the second order motions are processed by nonlinear system. 2009 Vol. 22 (3): 344-348 [Abstract] ( 327 ) [HTML 1KB] [ PDF 896KB] ( 732 ) 349 Semi-Supervised Proximal Support Vector Machine via Generalized Eigenvalues YANG Xu-Bing, PAN Zhi-Song, CHEN Song-Can A binary classifier, proximal support vector machine via generalized eigenvalues (GEPSVM), has been proposed recently. In this paper, with the characteristics of plane classifiers and manifold learning, an effective semi-supervised algorithm SemiGEPSVM is proposed. It keeps the performance of handling XOR problems and is suitable for more challenges, even with only one labeled sample per class. While the number of labeled samples is not satisfactory to generate plane, k-nearest neighbor is used to select the unlabelled samples. Otherwise, the proposed sample selection method with plane characteristics is adopted. Furthermore, it is proved that the proposed selection method is global optimization. And the experimental results of SemiGEPSVM are verified on one toy problem and some benchmark datasets. 2009 Vol. 22 (3): 349-353 [Abstract] ( 350 ) [HTML 1KB] [ PDF 484KB] ( 507 ) 354 Approximate Computation of Sensitivity of Adaline to Weight Perturbation WANG Bing-Hui, ZENG Xiao-Qin , ZHONG Shui-Ming The sensitivity of Adaline to weight perturbation is discussed. Considering the discrete feature of input and output of Adaline, the sensitivity is defined as the probability of an Adaline's output inversion due to the weight perturbation with respect to all possible inputs. By hypersphere model and analytical geometry technique, a method is proposed for approximately computing the sensitivity. Under the circumstance of high enough input dimension, the method is prior to other existing methods. It reduces computational complexity greatly with little sacrifice in precision and makes the sensitivity more practical. 2009 Vol. 22 (3): 354-359 [Abstract] ( 254 ) [HTML 1KB] [ PDF 556KB] ( 535 ) 360 Social Network Evolution Analysis Based on Graph Entropy GUO Rui, ZHONG Ning, LI Wen-Bin The order of networks in evolution is described by graph entropy. And the different evolution tendencies of connectivity entropy and centrality entropy are discussed. The different evolution tendencies reflect one characteristics of social networks inconsistency of multiscale. Then, the important nodes in networks are detected by entropy participation ratio. Finally, the reason of the distinct tendencies of global property and local property is shown in theory. 2009 Vol. 22 (3): 360-365 [Abstract] ( 348 ) [HTML 1KB] [ PDF 461KB] ( 577 ) 366 Extremum Decomposition Based Mixtures of Kernels and Its Improvement YE Qiao-Lin, YE Ning, ZHANG Xun-Hua In this paper, the property is used that any matrices can be decomposed into symmetric positive semi-definite ones using extremum decomposition. It is used on RBF kernel to generate a new kernel, called Ked. Combining Ked with global poly kernel, a mixed kernel with good classification performance is constructed. The classification experiments on UCI database are deployed with the mixed kernel. Compared with RBF kernel, the experimental results show that mixed kernel can decrease the number of support vectors and has better classification performance. Furthermore, it has good training time when the value of RBF kernel parameter is small. 2009 Vol. 22 (3): 366-373 [Abstract] ( 343 ) [HTML 1KB] [ PDF 495KB] ( 565 ) 374 Diversity Control Based on Distribution Entropy in Population-Based Search and Optimization XIN Bin, CHEN Jie, DOU Li-Hua, PENG Zhi-Hong A quantitative description of diversity in population-based search algorithms is put forward by comparing distribution entropy with variance. The problem of mode classification in individual space is presented for multimodal cases in optimization computation, and a classification method is proposed. On the basis of clustering analysis, the class distribution of individuals in search space is acquired. Furthermore, the diversity index described by distribution entropy is obtained. Then, diversity control is implemented by aggregation and dilation among individuals according to diversity. As an example, a first-order aggregation and dilation (A&D) algorithm for diversity control is presented and the setting of its parameters is analyzed. Simulation results demonstrate that the proposed algorithm performs better than the canonical genetic algorithm, the particle swarm optimization and the A&D search algorithm without classification. 2009 Vol. 22 (3): 374-380 [Abstract] ( 344 ) [HTML 1KB] [ PDF 382KB] ( 415 ) 381 Formal Model of Emotional Agent ZHANG Dong-Lei, SHI Zhong-Zhi, PAN Yu A formal model of emotional agent is presented to make an agent have the ability of handling emotions. Three sub-models, agent's mental state model, agent's mental state update model and the final emotional agent model, are defined progressively on three levels. By the theoretical update process axioms of several basic emotions, the model is proved to sufficiently embody emotional agent's essential characteristics and running mechanism. It has strong ability of representation and can be easily extended, which provide theoretical foundation for the construction of emotion rule base that supports emotion reasoning. The specific emotion update process and architecture of emotional agent are described. Simulation results show that the emotional agent can elicit emotions properly and make more intelligent decision, which proves the validity of the proposed model. 2009 Vol. 22 (3): 381-387 [Abstract] ( 303 ) [HTML 1KB] [ PDF 574KB] ( 463 ) 388 An Improved PCA Algorithm with Local Structure Preserving WANG Qing-Gang, LI Jian-Wei Locality preserving projection (LPP) is a local structure preserving method and the distances of neighboring points are minimized in the subspace of LPP. Combined with the geometric idea of LPP, an improved PCA with local structure preserving is proposed called locality preserving PCA (LP-PCA). By constructing the neighborhood graph and its complement, LP-PCA deals with the neighboring points and the far points distinguishingly. LP-PCA minimizes the distances between the neighboring points and simultaneously maximizes the distances between the far points. The improved algorithm can find the global structure of the high dimensional dataset with preserving its local structure. Some examples of the improved algorithm are given on toy datasets as well as on actual datasets. Experimental results show the effectiveness of LP-PCA. 2009 Vol. 22 (3): 388-392 [Abstract] ( 291 ) [HTML 1KB] [ PDF 365KB] ( 1154 ) 393 Attribute Reduction of Fuzzy Rough Sets Based on Variable Similar Degree ZHANG Hui-Zhe, WANG Jian, MEI Hong-Biao A fuzzy rough model based on variable similar degree is presented. It introduces the fuzzy similar degree and extends classical Pawlak's model. Firstly, the definitions of fuzzy similarity matrix and fuzzy difference degree matrix are given. Then, in the light of these definitions, the concepts of knowledge reduction are provided, such as attribute reduction, core, and algorithm of attribute reduction. Moreover, the relation between minimum reduction and core is proved. The proposed method obtains attribute reduction sets at different levels. Meanwhile, it keeps the classification accuracy with better flexibility by adjusting the similar precision. Compared to compact computational domain of fuzzy rough set, the proposed method has better classification accuracy and it provides a new idea for continued attribute data reduction. 2009 Vol. 22 (3): 393-399 [Abstract] ( 260 ) [HTML 1KB] [ PDF 388KB] ( 462 ) 400 Spatially Smooth and Complete Subspace Learning Algorithm LI Yong-Zhou, LUO Da-Yong, LIU Shao-Qiang A spatially smooth and complete subspace learning algorithm is proposed for feature extraction and recognition. Based on principle component analysis, spatially smooth subspace learning and locally sensitive discriminant analysis, the proposed algorithm preserves globally and locally geometrical structure and information of discrimination and spatial correlation. The globally geometrical features and locally spatial correlation information are extracted from original data samples, and then they are linearly transformed into new data samples. Subsequently, the best features are extracted for classification. Compared with general subspace learning algorithms, the proposed algorithm improves the recognition rate. Experimental results demonstrate the effectiveness of the proposed algorithm. 2009 Vol. 22 (3): 400-405 [Abstract] ( 279 ) [HTML 1KB] [ PDF 444KB] ( 383 ) 406 Spectral Clustering Algorithm of Image Texture Based on Rotated Complex Wavelet Transform XING Rui, XU Shu-Chang, ZHANG San-Yuan, ZHU Le-Qing As an important feature, texture plays a critical role in image retrieval. A clustering method is proposed based on image texture. Rotated complex wavelet (RCW) and dual-tree complex wavelet transform (DT-CWT) are used to decompose image into high frequency coefficients in twelve directions. The histogram signatures can be computed from each high frequency sub-band. Combined with other features, those signatures are employed to compute the similarity between data points for the improved spectral clustering to reduce dimensionality. In the final step, k-means is applied on the dimensionality-reduced data to get the clustering result. The proposed histogram signature for RCW and DT-CWT decomposition can capture the high frequency information in each direction effectively. In addition, an adaptive approach is proposed to compute the similarity between data points in spectral clustering. The experimental results show the proposed method outperforms the traditional methods remarkably. 2009 Vol. 22 (3): 406-410 [Abstract] ( 257 ) [HTML 1KB] [ PDF 446KB] ( 387 ) 411 TSK Fuzzy Systems Based on Fuzzy Partition and Support Vector Machines CAI Qian-Feng, HAO Zhi-Feng, LIU Wei An algorithm is presented to design a Takagi-Sugeno-Kang(TSK) fuzzy system with good generalization ability and robustness in high dimensional feature space by fuzzy clustering algorithms and support vector machines (SVM). Firstly, the antecedent membership functions are obtained by fuzzy clustering algorithms in the product space of the input variables. Then, the corresponding consequent parameters of the TSK model can be estimated from data using SVM. The kernel function of the proposed algorithm can be generated by the antecedent membership functions and it is proved to be a Mercer kernel. Finally, experimental results of three well-known datasets show that the proposed method has better generalization ability and robustness than the traditional techniques of TSK model and SVM. 2009 Vol. 22 (3): 411-416 [Abstract] ( 424 ) [HTML 1KB] [ PDF 394KB] ( 828 ) Researches and Applications 417 Speaker Verification Based on Adapted Gaussian Mixture Model Feature Mapping YANG Shi-Qing, DAI Bei-Qian, XU Min-Qiang, LIU Qing-Song To mitigate the channel effect of the handset speaker recognition system, a feature mapping (FM) method is proposed to eliminate the channel variability. Gaussian mixture model (GMM) is used to establish a channel-independent voice model, and the channel-dependent voice models are derived from the GMM using a well-known maximum a posteriori (MAP) adaptation algorithm. The difference of clustering gaussians describes the channel variability for different voice. The mismatch between train and test is compensated by mapping channel rules. Experimental results on NIST99 and 2004 SRE database show that the system performance can be increased by 14.7% and 15.18% by the proposed approach. 2009 Vol. 22 (3): 417-421 [Abstract] ( 337 ) [HTML 1KB] [ PDF 352KB] ( 730 ) 422 An Approach for Keyword-Based Image Retrieval in Unannotated Image Databases JIAO Jun, JIANG Yuan, LI Ming, ZHOU Zhi-Hua An approach is proposed for keyword-based image retrieval in unannotated image databases. After the keyword is input, the surrounding text information of the images is used to filter some irrelevant images, and then the visual information is extracted to select the relevant images. The data editing techniques are employed to refine the relevant images which are used as queries for retrieving images from the image databases. Experimental results show that the proposed method can achieve good retrieval performance in unannotated image databases. 2009 Vol. 22 (3): 422-426 [Abstract] ( 340 ) [HTML 1KB] [ PDF 369KB] ( 514 ) 427 Locally Linear Embedding and Its Improved Algorithm for Multi-Pose Ear Recognition XIE Zhao-Xia, MU Zhi-Chun, XIE Jian-Jun The current methods for ear recognition are discussed, and locally linear embedding (LLE) is employed to deal with multi-pose ear recognition. An improved locally linear embedding algorithm is presented. The improved LLE algorithm selects the neighbors according to the Hsim function, thus the instability of the neighbors in the higher dimensional space is well avoided. Experimental results demonstrate that LLE is feasible for multi-pose recognition and it has obvious advantages. Moreover, the improved LLE obtains a better recognition rate for multi-pose ear recognition. 2009 Vol. 22 (3): 427-432 [Abstract] ( 299 ) [HTML 1KB] [ PDF 552KB] ( 414 ) 433 Motion Activity Recognition Based on Abstract Hidden Markov Model QIAN Kun, MA Xu-Dong, DAI Xian-Zhong Recognition of human motion activity is essential in home-care robotic systems. In this paper, a probabilistic approach is proposed for human motion activity recognition based on the abstract hidden Markov model (AHMM). The AHMM is a well-suited hierarchical model for representing goal-directed motions at different levels of abstraction. In this model, the decision making process of agent is equivalent to an abstract Markov decision process (MDP). A model learning method is presented based on expectation-maximization algorithm to learn the observation model and the transition model respectively. Moreover, approximate inference of the AHMM is achieved by using Rao-blackwellised particle filters and thereby it enables efficient computation in recognizing motion patterns. Using trajectories derived from a visual tracking system, several indoor motion patterns are recognized. Experimental results validate the good performance of the proposed approach. 2009 Vol. 22 (3): 433-439 [Abstract] ( 456 ) [HTML 1KB] [ PDF 1052KB] ( 708 ) 440 Automatic Text Summarization Approach Based on Textual Unit Association Networks TAO Yu-Hui, ZHOU Shui-Geng, GUAN Ji-Hong An automatic text summarization approach is proposed based on textual unit association network. The word-based and sentence-based association networks are constructed respectively. For the word, a new approach is used to compute the word weights and then the weight of the sentence is evaluated based on the weights of words contained in the sentence. For the sentence, a new approach is presented to weight the salience of a sentence based on its cooccurrence information. Finally, salient sentences are extracted into the output summary till the desired summary length is satisfied. Experimental results show that the proposed approach can achieve better summarization performance than the existing methods. Moreover, the proposed scheme of term weighting can be used for keyword extraction, text classification and clustering and other information retrieval tasks. 2009 Vol. 22 (3): 440-444 [Abstract] ( 254 ) [HTML 1KB] [ PDF 312KB] ( 460 ) 445 Local Linear Space Partition Based Manifold Generation Algorithm CHEN Hua-Jie, PENG Dong-Liang A local linear space partition (LLSP) based manifold generation algorithm is proposed to extend the mapping function obtained by manifold learning of new data points. Using the dimension-fixed projection distance (DFPD), the dimension-fixed projection vector quantization (DFPVQ) algorithm is presented to cover the whole manifold with several local linear spaces (LLS). Then, the simplified linear mapping functions are constructed in LLS. As for new data point, the corresponding LLS is found and the mapping value in low dimension is estimated by the simplified linear mapping function. The superiority of the proposed algorithm is confirmed by experimental results both on synthesized data and handwritten digits image dataset. 2009 Vol. 22 (3): 445-451 [Abstract] ( 315 ) [HTML 1KB] [ PDF 1000KB] ( 541 ) 452 An Alopex Based Evolutionary Optimization Algorithm LI Shao-Jun An Alopex based evolutionary algorithm is proposed. Its salient feature is randomly selecting two individuals and computing their objective values. According to the information of the two individuals, the probability of search direction is ascertained. By iterative computing, the global optimum is obtained. It has the advantages of both gradient methods and simulation anneal algorithm to some extent. The anneal temperature is self-adjusting over the proceeding of evolution. The proposed algorithm is used to optimize the benchmark functions and the kinetic parameters of 2-chlorophenol oxidation in supercritical water. The experimental results demonstrate that the proposed algorithm is superior to the original evolutionary algorithms, especially for the multi-apices function problems. 2009 Vol. 22 (3): 452-456 [Abstract] ( 377 ) [HTML 1KB] [ PDF 327KB] ( 410 ) 457 A Clustering Algorithm for Network Objects with Direction Factors TANG Liang, FANG Ting-Jian Clustering methods are analyzed in which Euclidean distance and network distance are used as a similarity measure respectively. The neighbor correlation between objects on a spatial network is discussed and a clustering algorithm is proposed for network objects with consideration of direction factors. The algorithm combines the two distances as the similarity measure of clustering by using the neighbor correlation. The analysis and experimental results indicate that the effectiveness of the proposed algorithm is better than those only using one measure. 2009 Vol. 22 (3): 457-462 [Abstract] ( 227 ) [HTML 1KB] [ PDF 414KB] ( 460 ) 463 Bilayer Video Segmentation Based on Random Ferns CHU Yi-Ping, CHEN Qin, HUANG Ye-Jue, ZHENG He-Rong A random ferns based method is proposed for bilayer video segmentation with the capability of segmenting monocular video automatically. Motion feature dictionary is constructed by clustering the motion features of the video, and the motion features are modeled by random ferns. The video colors, motion features and neighboring relationships are constrained by using conditional random fields. The graph-cut algorithm is adopted for solving globally optimal segmentation results. The experimental results demonstrate the validity of the proposed algorithm, and the results of the proposed method are compared with other algorithms on different video data. 2009 Vol. 22 (3): 463-467 [Abstract] ( 300 ) [HTML 1KB] [ PDF 604KB] ( 503 ) 468 Palmprint Recognition Based on Probability Template and Fuzzy Logic ZHU Le-Qing, ZHANG San-Yuan, ZHANG Yin, YE Xiu-Zi A palmprint recognition algorithm is proposed based on principal palm lines. According to the previous knowledge, an irregular geometrical shape is used to get the valid region to decrease the interference of bulky noises. The information about the positions and structures of palm lines and the intensity are included in the final extracted principal lines. Thus, sufficient clues are provided for palmprint recognition. A conception of probability template of principal lines is presented, which suppresses the influence of random noises. Features of several training samples are merged into one template to guarantee the integrity of features and improve the matching efficiency. Fuzzy logic is adopted in the matching algorithm. Experimental results show that the palmprints can be verified and distinguished with high precision by the proposed method. 2009 Vol. 22 (3): 468-474 [Abstract] ( 327 ) [HTML 1KB] [ PDF 1328KB] ( 531 ) 475 Adaboost Object Tracking Algorithm JIA Jing-Ping, ZHANG Fei-Zhou, CHAI Yan-Mei An Adaboost based algorithm for object tracking in image sequences is proposed. In this algorithm, tracking is considered as a binary classification problem. Firstly, the linear combination of R, G, and B with integer coefficients is used to generate the candidate features. Features are selected for the design of weak classifiers according to the two-class variance ratio. Then, a strong classifier is built on the weak classifiers. For each incoming frame, a likelihood image of the object is created according to the classification results of pixels by the strong classifier. The trust region method and the scale space theory are employed to locate the blobs in the likelihood image, and thus the object tracking is fulfilled. The changes of illumination often cause the changes of features. The adaptability of the proposed algorithm is improved by online integration of new weak classifiers and automated weights update of the used ones. Based on the tracking results of sequence examples, the proposed algorithm can adapt to feature changes, track object in cluttered background and describe the object accurately with better tracking precision. 2009 Vol. 22 (3): 475-480 [Abstract] ( 372 ) [HTML 1KB] [ PDF 1796KB] ( 1099 ) 481 Universal Approximation of Fuzzy Functions by Polygonal Fuzzy Neural Networks with General Inputs HE Chun-Mei, YE You-Pei, LI Jian, XU Wei-Hong Firstly, a class of feedforward fuzzy neural networks (FNNs), polygonal FNNs, is proposed based on a redefined extension principle and fuzzy arithmetic.Then, while the inputs are general fuzzy numbers and the active functions are monotone continuous sigmoid functions, the topologic structure and the related properties of the polygonal FNNs are analyzed systemically. Some theorems for the continuous fuzzy function can be approximated to any degree of accuracy by polygonal FNN and they are proved. Finally, the equivalent conditions are presented. Thus the problem whether the polygonal FNNs with general inputting fuzzy numbers is the universal approximator to the class of continuously increasing fuzzy function is solved, and consequently the application areas of polygonal fuzzy neural networks are extended. 2009 Vol. 22 (3): 481-487 [Abstract] ( 281 ) [HTML 1KB] [ PDF 350KB] ( 439 ) 488 Scanned English Document Retrieval Based on OCR and Word Shape Coding XIA Yong, DAI Ru-Wei, XIAO Bai-Hua, WANG Chun-Heng Two commonly used methods for scanned document retrieval are analyzed, namely retrieval based on optical character recognition (OCR) and retrieval based on word shape coding. A new strategy of combining these two methods based on recognition confidence is given. Furthermore, a new way for word shape coding based on typographic feature and stroke is presented and it is tolerant to fonts. Experiments are conducted based on different word indexing and the results verify the validity of the proposed method. 2009 Vol. 22 (3): 488-493 [Abstract] ( 301 ) [HTML 1KB] [ PDF 456KB] ( 533 ) 494 Accuracy Analysis of SVM Based Ballistic Recognition Approach TAO Qing, NA Jian, FENG Yong, LIU Xin The ballistic recognition accuracy is further discussed and analyzed in terms of simulation experiments. Firstly, the reason for causing the misclassified samples is analyzed. Then, the influences of the number of training samples, the interval of sampling and the noise of radar on the recognition accuracy are respectively discussed in detail. Finally, several significant and interesting results are achieved. 2009 Vol. 22 (3): 494-498 [Abstract] ( 383 ) [HTML 1KB] [ PDF 346KB] ( 774 ) 499 Application of Higher-Order Statistics Features of ICA Coefficients in Texture Classification XU Xiao-Hong, YANG Xue-Zhi, YANG De-Mei, Gao Jun ICA coefficients are non-Gaussian in independent component analysis model. The high-order statistical features are used in characterization of non-Gaussian feature. The combined moments of variance, skewness and kurtosis are proposed to describe the ICA coefficients probability distributing characteristic. The combined moments are used in texture classification and it can achieve better classification performance than the previously reported ICA features. Furthermore, L-moments are used to improve robustness in moments estimation and to get better performance than the ordinary moments. 2009 Vol. 22 (3): 499-505 [Abstract] ( 255 ) [HTML 1KB] [ PDF 371KB] ( 529 )

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