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....

2008 Vol.21 Issue.3, Published 2008-06-01 Papers and Reports    Researches and Applications    urveys and Reviews        Papers and Reports 273 Smoothing Functions for Support Vector Regressions XIONG Jin-Zhi, HU Jin-Lian, YUAN Hua-Qiang, HU Tian-Ming, PENG Hong Smoothing functions can transform the unsmooth support vector machines (SVMs) into smooth ones, and thus better regression results are generated. A smoothing function was used by Lee et al. to approximate the square ofε-insensitive loss function, therefore the ε-insensitive smooth support vector regression (ε-SSVR) was proposed. In this paper, using techniques of interpolation function and function composition, a kind of smoothing functions is proposed for ε-insensitive support vector regressions (SVRs). Smooth approximations of the plus function are firstly derived and then applied to approximate the square of the ε-insensitive loss function. Theoretical analysis shows that the approximation accuracy of the proposed smoothing functions is an order of magnitude higher than that of the existing ones. Better regression results are yielded and the new kind of smoothing functions is provided for SVRs. 2008 Vol. 21 (3): 273-279 [Abstract] ( 327 ) [HTML 1KB] [ PDF 365KB] ( 485 ) 280 Statistical Acoustic Model Based Unit Selection Algorithm for Speech Synthesis LING Zhen-Hua, WANG Ren-Hua A statistical acoustic model based unit selection algorithm for speech synthesis is proposed. During training stage, the acoustic models for contextual dependent phonemes are built up by using acoustic features extracted from the training data, such as spectral parameters, F0, and segmental and prosodic labels in the corpus. The hidden Markov model (HMM) is adopted as the model structure. During synthesis stage, the optimal phoneme unit sequence is searched in the speech corpus by maximizing the probabilistic likelihood between its acoustic features and the sentence HMM constructed with the contextual information of input text. Finally, the waveforms of the selected candidate units are concatenated and smoothed to produce the synthesized speech. Based on the proposed method, a Chinese speech synthesis system using initials and finals as the basic concatenation units is constructed. Results of listening test prove that the proposed method can achieve better naturalness of synthesized speech compared to the conventional method. 2008 Vol. 21 (3): 280-284 [Abstract] ( 335 ) [HTML 1KB] [ PDF 408KB] ( 1325 ) 285 An Improved Particle Swarm Optimization Algorithm for Vector Quantization LI Xiao-Jie, XU Lu-Ping, YANG Li An improved particle swarm optimization algorithm for vector quantization is proposed. The Concept of comprehensive learning in comprehensive learning particle swarm optimization (CLPSO) is adopted and merged into the learning strategies of original particle swarm optimization (PSO). The mapping between a particle and its example particle is built. And the particle can learn from the mapped dimensions in the example particle instead of the corresponding dimensions. Thus, the local search ability is greatly enhanced as well as the diversity of the swarm is effectively maintained. The experimental results show that the algorithm can effectively alleviate the problem of premature convergence and obtain good reconstruction image quality. 2008 Vol. 21 (3): 285-289 [Abstract] ( 295 ) [HTML 1KB] [ PDF 440KB] ( 487 ) 290 A Robust Computer Real-Time Accompaniment System LUO Li, WANG Zeng-Fu Computer automatic accompaniment is studied from system viewpoint, and a computer system providing real-time accompaniment for electronic instrument is developed. The system includes two main modules: score-following and real-time accompaniment. To solve score-following problem, a dynamic programming algorithm is proposed based on extended window and reconstruction of matcher. The algorithm can locate input musical notes online in the score when the musical performance is full of errors compared to the score. To improve the quality of real-time accompaniment music, the bar is inducted as the unit of accompaniment and an accompaniment strategy is presented which determines how to add accompaniment online according to the location results. Experiment results demonstrate that the system has high performance in error resistance and real-time accompaniment expression. 2008 Vol. 21 (3): 290-297 [Abstract] ( 271 ) [HTML 1KB] [ PDF 819KB] ( 543 ) 298 An Image Denoising Algorithm Based on Noise Detection JIN LiangHai, LI DeHua An efficient image denoising algorithm is introduced. Firstly, image pixels are classified into noisy pixels and noise-free pixels by four directional operators. Then an adaptive weighted median filter is designed to remove and restore the detected noisy pixels and keep the noise-free ones unchanged. Experimental results indicate that the proposed algorithm preserves image details well while removing impulsive noise efficiently, and its filtering performance is significantly superior to the classical median filter and some other typical and recently developed improved median filters. 2008 Vol. 21 (3): 298-302 [Abstract] ( 309 ) [HTML 1KB] [ PDF 826KB] ( 677 ) 303 Clonal Selection Optimization for Multi-Mode Resource Constrained Project Scheduling Problem PAN Xiao-Ying, LIU Fang, JIAO Li-Cheng Based on the analysis of the characteristics of project optimization scheduling, a clonal selection algorithm for multi-mode resource constrained project scheduling problem (CSA-MRCPSP) is proposed. It is used to obtain the optimal scheduling sequences so that the duration of the project is minimized. Some strategies are adopted such as schedule encoding, semi-random initialization, and restricted mutation operator. CSA-MRCPSP synthesizes the characteristics of project scheduling global search, diversity, and no prone to premature in immune clonal selection. Thus the cost is reduced with the optimal solution being found. The experimental results on PSPLEB show CSA-MRCPSP has good performance and it can find optimal solution in reasonable time for most instances. Furthermore, compared with other heuristic methods, CSA-MRCPSP also has some advantages, including higher optimal proportion and lower average deviation. 2008 Vol. 21 (3): 303-309 [Abstract] ( 232 ) [HTML 1KB] [ PDF 455KB] ( 479 ) 310 PSO-Based Self-Tuning Control for Time-Varying Delay and Variable Structure System LIN Wei-Xing, OU Chao , LIU Peter X., LI Wen-Lei A design method is presented for the self-tuning control algorithm in a time-varying delay and variable structure system. A self-tuning regulator is optimized by the particle swarm optimization (PSO) algorithm, combined with generalized minimum variance. Using an implicit identification, the method can track the errors of the system to increase the precision and decrease the computational burden. It is adaptable for the typical industrial process control, especially for time-varying and large time-delay models. Results of simulation and comparison show its advantages in robust and tracing precision. 2008 Vol. 21 (3): 310-316 [Abstract] ( 284 ) [HTML 1KB] [ PDF 555KB] ( 430 ) 317 Online Freehand Sketching Recognition Using Fuzzy Theory WANG Shu-Xia, GAO Man-Tun, QI Le-Hua Based on geometric features and hybrid features, the fuzzy recognition methods of online freehand sketching are discussed. The fuzzy features and corresponding extraction methods of geometric primitives are introduced, and a general fuzzy membership function is designed by analyzing the triangle membership function. The possibilities of geometric primitives which the stroke belongs to are obtained through numerous experiments. The system supplies the human-computer interactive determinant of the stroke through cutting the value of membership function. If necessary, the designer can help the system make the proper inference when either wrong choice is made by the system, or plural possibilities are analyzed. By the methods, designers' intentions can be inferred and input sketches can be interpreted into more exact 2D geometric primitives, including straight line, polyline, circle, circular arc, ellipse, elliptical arc, hyperbola and parabola. The effectiveness of the algorithm is demonstrated preliminarily by experiments. It lays the foundation for 3D object recovering and conceptual design of freehand sketching. 2008 Vol. 21 (3): 317-325 [Abstract] ( 369 ) [HTML 1KB] [ PDF 510KB] ( 646 ) 326 Face Recognition Based on Symmetrical 2DPCA YANG WanKou, REN MingWu, YANG JingYu An algorithm is proposed, called symmetrical two dimensional principal component analysis (S2DPCA). It is based on the theory of function decomposition in algebra and mirror symmetry in geometry. Firstly, mirror transform is applied to images. Then, the images are decomposed into even and odd symmetrical images, and 2DPCA are performed on the even and odd images respectively. According to the idea of selective ensemble, the more discriminant eigenvectors of the even and odd image space are selected to construct the final eigenspace. Finally, the even and odd 2DPCA features are gotten by projecting samples onto the eigenspace. Both theoretical analysis and experimental results demonstrate that the algorithm can enlarge the number of training samples and raise the recognition rate. 2008 Vol. 21 (3): 326-331 [Abstract] ( 244 ) [HTML 1KB] [ PDF 383KB] ( 636 ) 332 Fuzzy Support Vector Machine Method Based on Border Vector Extraction WU Qing, LIU San-Yang, DU Zhe A fuzzy support vector machine (SVM) based on border vector extraction is presented. It overcomes the disadvantage of the sensitivity to noises and the outliers in the training samples. Border vectors, which are possible support vectors, are selected as new samples to train SVMs. The number of training samples is reduced and thus the training speed is improved. The fuzzy membership is defined according to the distance from border vectors and outliers to their hypersphere centers. Consequently the effect of noises and outliers is weakened and support vectors are improved to design a classifier. Experimental results show that by the proposed method the machine is less sensitive to noises and outliers than by the traditional SVMs and the fuzzy SVMs based on the distance between a sample and its cluster center. Furthermore, the proposed method has better generalization ability and higher learning speed than the others. 2008 Vol. 21 (3): 332-337 [Abstract] ( 275 ) [HTML 1KB] [ PDF 397KB] ( 666 ) 338 A Multi-Cluster Structure Based Gaussian Dynamic Particle Swarm Optimization Algorithm NI Qing-Jian, XING Han-Cheng, ZHANG Zhi-Zheng, WANG Zhen-Zhen The method of population generation in Gaussian dynamic particle swarm optimization algorithm (GDPSO) is analyzed detailedly. Aiming at the problem of premature convergence of Gbest version and the slow search speed of Lbest version in original particle swarm optimization, a novel neighborhood topology structure called multi-cluster structure is proposed. In the proposed population structure, particles in one cluster share the information with each other, and clusters exchange their experiences through loose connection between particles. Thus, neighborhood topology is designed to coordinate exploration and exploitation. GDPSO, with several population topologies including the multi-cluster structure, is tested on four benchmark functions which are commonly used in the evolutionary computation. Experimental results show that the GDPSO with the proposed neighborhood topology can significantly speed up the convergence and efficiently improve the global search ability. 2008 Vol. 21 (3): 338-345 [Abstract] ( 290 ) [HTML 1KB] [ PDF 430KB] ( 443 ) 346 Relative Division of Overlapping Space Based Biomimetic Pattern Recognition WU Yan, YAO Xiao, WANG Shou-Jue To realize the effective coverage in feature space, a method of relative division of overlapping space based biomimetic pattern recognition (RDBPR) is proposed in high dimensional space. It can relatively classify the samples on the basis of cognition. In the overlapping space caused by big threshold value, the sample is classified by calculating the distance to the relative subspace, and thereby the correct recognition rate can be improved with low misclassification rate. The experimental results of face recognition show RDBPR has higher correct recognition and better stability than the traditional classifiers. 2008 Vol. 21 (3): 346-350 [Abstract] ( 331 ) [HTML 1KB] [ PDF 338KB] ( 525 ) urveys and Reviews 351 A Survey of Stochastic Diffusion Search WANG Li-Fang, ZENG Jian-Chao As one of swarm intelligence optimization algorithms, the stochastic diffusion search is characterized by partial function evaluation and one-to-one recruitment mechanism. These characteristics make the algorithm high computation efficiency and robustness of the stochastic diffusion search.Based on the survey of basic principles and the research actuality of stochastic diffusion search, the existing problem and features are analyzed, and some future research directions about the stochastic diffusion search are delineated. 2008 Vol. 21 (3): 351-356 [Abstract] ( 309 ) [HTML 1KB] [ PDF 378KB] ( 451 ) Researches and Applications 357 A Reinforcement Learning Based ART2 Neural Network: RL-ART2 FAN Jian, FEI Mei-Rui A reinforcement learning based ART2 neural network (RLART2) is proposed and its learning algorithm is given. It is capable of online learning without training samples by using the characteristic of alteration with environment of reinforcement learning. In RLART2, the output classified pattern is got by inner competition of ART2, then the running effect of the classified pattern is gained and evaluated through altering with environment. With enough time of being online and interactive learning with environment, a certain recognition ratio of ART2 neural network is attained. The simulation results of path planning for mobile robot indicate that the collision times of robot is effectively decreased by using RLART2 . Moreover, the rationality and validity of RLART2 are also demonstrated by the results. 2008 Vol. 21 (3): 357-362 [Abstract] ( 374 ) [HTML 1KB] [ PDF 496KB] ( 457 ) 363 An HMM Based Normality Processing Method without Overflow for High Dimensional Sample Set TANG Jing-Hai, ZHANG You-Wei Aiming at the overflow of the hidden Markov model (HMM) observation probability, a method is proposed, called Normality Processing. Firstly, the chi-square plot is used to test normality of the sample set, the transformation of square root is performed. The feasibility of the proposed method is validated on the expression sequences database of CED-WYU(1.0) and Cohn-Kanade (CMU). The person-independent expression recognition experiment is made with continuous HMM based on the optical flow features and a better result is obtained when the normality processing is used. 2008 Vol. 21 (3): 363-368 [Abstract] ( 269 ) [HTML 1KB] [ PDF 414KB] ( 423 ) 369 Integrated Extraction of Handwritten Numeral Strings in Form DocumentBased on Hybrid Binarization ZHENG Tian-Xiang, XIE Liang, YANG Li-Hua The handwritten numeral string extraction in form document is studied. A method is proposed to effectively discern and capture the characters from overlapping borders based on hybrid binarization. Two key problems are investigated in detail including the location and the extraction on the cell of interest (COI) with broken strokes mended. The extracted handwritten characters remain integrated even for characters in different writing styles. Experimental results demonstrate that the proposed method is efficient. 2008 Vol. 21 (3): 369-375 [Abstract] ( 272 ) [HTML 1KB] [ PDF 612KB] ( 589 ) 376 Unsupervised SAR Image SegmentationBased on Immune K-Means Clustering BO Hua, MA Fu-Long, JIAO Li-Cheng Combined with the information entropy characteristic of image texture and the co-occurrence-matrix concept, a practical unsupervised SAR image segmentation algorithm is presented based on immune K-means clustering. It overcomes the disadvantages of local optima and sensitivity to the values and noises, and has the same fast-convergence advantage as K-means method. The theoretical analysis and experimental results show that the proposed algorithm has low computing complexity and strong robustness. 2008 Vol. 21 (3): 376-380 [Abstract] ( 323 ) [HTML 1KB] [ PDF 631KB] ( 427 ) 381 A Visual Memory Model with Amnesic Function and Its Application in Attention Selection GUO Chen-Lei, ZHANG Li-Ming Based on the previous attention selection model with visual memory as top-down guidance, a visual memory model is put forward with online learning and forgetting, called amnesic incremental hierarchical discriminant regression (AIHDR) tree, to mimic human short-term memory (STM) and long-term memory (LTM). A self-supervised competition neural network (SSCNN) combines the information from both bottom-up and top-down to find out the focus of attention (FoA). The connection weights in SSCNN can be updated in real-time according to the environment. Experimental results show that the proposed model can mimic the shift of human attention and stare at an interesting object consciously when environment changes. 2008 Vol. 21 (3): 381-387 [Abstract] ( 264 ) [HTML 1KB] [ PDF 1012KB] ( 726 ) 388 A Kernel Based Supervised Manifold Learning Algorithm LI Jun-Bao, PAN Jeng-Shyang A kernel based supervised manifold learning method is presented to solve the problems on parameter selection with locality preserving projection and inability in nonlinear feature extraction, which is unresolved by the currently proposed manifold learning algorithm. The proposed algorithm is an improvement of locality preserving projection (LPP). The nearest neighbor graph is created with the class label information of the samples, and nonparametric similarity measure is used. The kernel method is used to solve the limitation problem of the nonlinear separability for LPP. The feasibility and effectivity of the proposed algorithm are testified on two databases. 2008 Vol. 21 (3): 388-393 [Abstract] ( 225 ) [HTML 1KB] [ PDF 324KB] ( 457 ) 394 Improvements on Search Process and Search Subspace in Active Shape Model HE Liang-Hua, HU Die, JIANG Chang-Jun The method of active shape model (ASM) is quite commonly used for alignment in recent years. However, its search subspace has a number of limitations on reconstructing the changeable shape in real life. In addition, since there is no restriction during searching, the result is unstable. In this paper, some improvements on search subspace and search process are proposed. By adding eigen-shape variance information to search subspace, the new search subspace can reconstruct the shape more generally. In the meantime, the search errors are analyzed and searching information of the next step is obtained. The whole search processing is iterated and connected with the feedback of search error. The feedback and iteration make the search processing more active and the final search result is unique. Several face alignment experiments are designed and the results show that the proposed method improves alignment precision greatly. 2008 Vol. 21 (3): 394-400 [Abstract] ( 255 ) [HTML 1KB] [ PDF 981KB] ( 583 ) 401 Method for Similar Pattern Discovery in Time SeriesBased on Neural Network ZHANG Peng,ZHANG Jian-Ye,DU Jun,LI Xue-Ren According to the unsupervised neural network theory of clustering, a method is proposed for similar pattern discovery in time series database. Aiming at the poor capability of the neural network for handling the time change process sequence, the original data are mapped into the feature pattern space by means of fast discrete cosine transform (FDCT) for dimension reduction. The advantages of artificial neural network as similarity measurement model are analyzed and the range query algorithm is presented. The simulation results show that the proposed algorithm has the property of multi-scale, and compared with Euclidean distance and Slop distance, it can reflect similarities of time series under various resolutions. 2008 Vol. 21 (3): 401-405 [Abstract] ( 347 ) [HTML 1KB] [ PDF 423KB] ( 743 ) 406 An Algorithm for Adaptive Neighborhood Selection WEI Lai, WANG Shou-Jue, XU Fei-Fei An automatic neighborhood selection algorithm for manifold learning is proposed. It is suitable for all the manifold learning algorithms which need select the neighbors to get locally linear information. Through the algorithm, the proper neighborhood size of a dataset can be determined even under different data density and curvature. By adopting this method, the locally multidimensional scaling can reduce the dimensionality of data based on the suitable neighborhood, and the low-dimensional representations of the data can be get through global alignment. The experiment shows the algorithm can recover the sophisticated geometry structure of the data sets. 2008 Vol. 21 (3): 406-409 [Abstract] ( 259 ) [HTML 1KB] [ PDF 467KB] ( 502 ) 410 A Class-Information-Incorporated Kernel Principal Component Analysis Method LI Yong-Zhi, YANG Jing-Yu, WU Song-Song A supervised feature extraction method based on kernel principal component analysis (KPCA) is presented. In feature extraction the class information of the training kernel sample is sufficiently utilized, and the simple mathematical formulation is employed which is similar to KPCA. Thus, the method is named as class-information-incorporated kernel principal component analysis (CIKPCA). Furthermore, a new classification strategy is presented by fusing two kinds of feature vectors to improve the recognition rate. The experimental results on three databases show that the proposed method is better than KPCA in terms of recognition rate, and it even outperforms KLDA. 2008 Vol. 21 (3): 410-416 [Abstract] ( 307 ) [HTML 1KB] [ PDF 656KB] ( 429 ) 417 An Improved Sequential IB Algorithm for Document Clustering YE Yang-Dong, ZHANG Jie, LIU Dong To solve the problems of local optima and low efficiency in sequential information bottleneck (sIB) algorithm for document clustering, an improved sIB algorithm is proposed, namely SA-isIB. By a reasonable annealing sequence, a certain proportional of documents are selected randomly from the initial clustering solution of basic sIB algorithm. Then the clustering labels of selected documents are revised and the solution is optimized iteratively. After the process of simulated annealing, higher accuracy document clustering solutions are obtained. Experimental results on document datasets show that by using SA-isIB algorithm the accuracy of sIB algorithm for document clustering is improved efficiently. 2008 Vol. 21 (3): 417-422 [Abstract] ( 299 ) [HTML 1KB] [ PDF 473KB] ( 661 )

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