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....

2011 Vol.24 Issue.4, Published 2011-08-25 Orignal Article        Orignal Article 457 An Improved Cross-Language Model Adaptation Method for Speech Synthesis LIU Hang, LING Zhen-Hua, GUO Wu, DAI Li-Rong Cross-language model adaptation in statistical parametric speech synthesis is used for rapidly constructing a text-to-speech (TTS) system with the target speakers characteristics when the source and the target speakers languages are different. In this paper, the conventional cross-language adaptation method based on phone-mapping and triphone models is improved by two means. Firstly, phone mapping combined with data-selection is adopted to improve its reliability. Secondly, cross-language prosodic information mapping is introduced to make use of prosodic information, which is ignored in the triphone model. Experiments on Chinese-to-English adaptation show that the synthesized speech using the improved method has much better naturalness and speaker similarity compared with the result of conventional method. 2011 Vol. 24 (4): 457-463 [Abstract] ( 384 ) [HTML 1KB] [ PDF 524KB] ( 633 ) 464 3D Robust Gait Recognition Based on Manifold Learning LIU Hai-Tao, WANG Zeng-Fu, CAO Yang Aiming at the situation that many approaches for gait recognition are based on a single camera, an approach of gait recognition based on stereo vision is proposed. Firstly, 3D coordinates of human body contour are gotten by stereo matching. Then, 3D body contour descriptor (3D-BCD) is constructed to get the gait feature of human. The noise and glitch are eliminated by noise-eliminated method. Thus, manifold learning (Laplacian Eigenmaps) is used for dimensionality reduction. The nearest neighbor classifier (NN) and the nearest neighbor classifier about template (TNN) are used for classifying category. Finally, a series of experiment results on the stereo gait database of PRLABⅡ and the irregular test stereo gait dataset ExN proved out the high correct classification rate and the strong robustness of the proposed approach. And the approach is not related with the distance between camera and the walking path. Moreover, it has stronger robustness with the incomplete gait sequences, the changes of human behavior, the changes of apparels, and carrying a bag. 2011 Vol. 24 (4): 464-472 [Abstract] ( 375 ) [HTML 1KB] [ PDF 698KB] ( 786 ) 473 Based on Locality Regularized Generalization Error Bound XUE Hui, CHEN Song-Can Feature selection is a hot topic in current pattern recognition. Filter and wrapper approaches are two of the most important policies to evaluate feature subsets in feature selection algorithms. However, they both can not guarantee the generalization performance of the following designed classifier. To solve these problems in the two approaches, a locality regularized generalization error bound is firstly introduced which embeds the manifold structure information hidden in the input samples. Furthermore, a hybrid filter-wrapper feature selection algorithm is proposed, which uses the locality regularized generalization error bound as the evaluation function as well as the locality regularization method as the classifier. As a result, the proposed algorithm can not only keep high computational efficiency, but also guarantee the good generalization performance of the following classifier. Experimental results validate the superiority of the algorithm. 2011 Vol. 24 (4): 473-478 [Abstract] ( 380 ) [HTML 1KB] [ PDF 413KB] ( 638 ) 479 Multi-Lateral Multi-Issue Negotiation Based on Adaptive Differential Evolution Algorithm BI Xiao-Jun, XIAO Jing To improve efficiency and stability of the negotiation multi-agent based e-commerce, an adaptive differential evolution (ADE) algorithm is proposed and applied to the multi-lateral multi-issue simultaneous bidding negotiation. The differential evolution (DE) is one of the current best evolutionary algorithm for global optimization over continuous spaces. The comparative experimetal results show that the adaptive DE algorithm can gain the optimal negotiation result more efficiently and more stably than the hybrid genetic algorithm (HGA) in multi-literal multi-issue negotiation. 2011 Vol. 24 (4): 479-483 [Abstract] ( 320 ) [HTML 1KB] [ PDF 367KB] ( 598 ) 484 Extracting Attack to DCT Domain Sequential LSB Stegangraphy CHEN Jia-Yong, ZHU Yue-Fei, ZHANG Wei-Ming, LIU Jiu-Fen A stego-only extracting attack method is proposed based on a spatial domain feature. The discontinuity of JPEG image coding blocks in spatial domain is used to construct a time series. The problem of extracting attack is changed to the problem of estimating the change-points in time series. Then, a technique for order preference by similarity to ideal solution (TOPSIS) model is built to estimate the change-points. The model can be applied to not only sequential JSteg but also an improved sequential JSteg algorithm which resists to Chi-square test attack. The experimental result indicates that average error for estimating the message embedding start point and end point is about 260 coefficients. 2011 Vol. 24 (4): 484-491 [Abstract] ( 560 ) [HTML 1KB] [ PDF 496KB] ( 645 ) 492 A Feature Extraction and Recognition Approach for Accelerometer Based Virtual Handwriting Digit XUE Yang, JIN Lian-Wen An approach for accelerometer based virtual digit feature extraction and recognition is proposed. Firstly, accelerations are projected on three planes respectively. Secondly, the rotation feature of the handwriting is extracted and coded. Then, normalized edit distance is used to measure the difference between different rotation feature codes. Finally, based on the rotation feature and edit distance, an algorithm of virtual handwriting digit recognition is given. Compared with time-domain feature, peak-valley feature and FFT feature, the proposed approach is effective. 2011 Vol. 24 (4): 492-500 [Abstract] ( 421 ) [HTML 1KB] [ PDF 573KB] ( 903 ) 501 A New Color Space YCH with Strong Clustering Power for Face Detection ZHONG Zhi-Guang Using information of skin color to detect face region is quick and effective. However, it is very difficult to choose a suitable color space. In this paper, A novel adaptive color space YCH is proposed. It fuses the merits of the commonly used color spaces into a new simple nonlinear transformation. All the transformation coefficients can adjust automatically according to the respective characteristics of each pixel in the face image, and then eliminate effectively various unfavorable influence factors to the classifying result of the skin color and non-skin color. Experimental results demonstrate the good discriminating power of the proposed color space to all kinds of face images. 2011 Vol. 24 (4): 501-505 [Abstract] ( 336 ) [HTML 1KB] [ PDF 316KB] ( 600 ) 506 An Improved Algorithm for Interactive Dynamic Influence Diagrams LI Bo, LUO Jian, YIN Hua-Yi, TIAN Le Interactive Dynamic Influence Diagrams(I-DIDs), as graphic models based on probabilistic graphical theory, are proposed to represent, the sequential decision-making problem over multiple time steps in the presence of other interacting agents. The algorithms for solving I-DIDs are haunted by the challenge of an exponentially growing space of candidate models ascribed to other agents over time. In this paper, in order to reduce the candidate model space according the behaviorally equivalent theory, a more efficient way to construct Epsilon behavior equivalence classes is discussed that using belief-behavior graph (BBG). A method of solving I-DIDs approximately is presented, which avoids solving all candidate models by clustering models with beliefs that are spatially close and selecting a representative one from each cluster. The simulation results show the validity of the improved algorithm. 2011 Vol. 24 (4): 506-513 [Abstract] ( 332 ) [HTML 1KB] [ PDF 527KB] ( 588 ) 514 Object Scale Perception Strategy with Embedding Entropy Region Manifold WU Ke-Wei, XIE Zhao, GAO Jun This paper mainly focuses on the issue of generic object scale perception. A computational model in entropy-domain space is presented for scene object description to pursue the underlying entropy manifold in statistical way. The corresponding algorithm approximately follows perceptual hierarchy in human-vision biologically via quad-tree pyramid structure, which can automatically choose the appropriate scale of various objects via proposed scale evaluation function. The sufficient experiments truly demonstrate the effective scale description in entropy region manifold with proper location, and provide additional priori information for object scale perception. 2011 Vol. 24 (4): 514-520 [Abstract] ( 336 ) [HTML 1KB] [ PDF 471KB] ( 680 ) 521 A Proof for Minimal Game Trees Leaf Node Number Theorem ZHANG Ming-Liang, WU Jun, LI Fan-Zhang A concise proof for minimal game trees leaf node number theorem is presented according to some deficiencies in its pervious proofs, and also some misunderstandings of minimal game-tree are clarified. On the analyses and experiments of the efficiency source of the window searches, this paper reveals the fact that the improvement of the window searches efficiency results from the position of the window. This qualitative conclusion, which contains some inconsistencies from the common knowledge, gives accurate comprehension and utilization of window searches. 2011 Vol. 24 (4): 521-526 [Abstract] ( 280 ) [HTML 1KB] [ PDF 466KB] ( 735 ) 527 Classification in Networked Data: A Survey XIONG Wei, ZHOU Shui-Geng, GUAN Ji-Hong The rapid increase of network applications generates a lot of networked data. Classification in networked data is recently an important research issue of data mining field. The state of the art techniques of classification in networked data is surveyed. Firstly, the basic concepts of classification in networked data are introduced. Then, major classification algorithms of networked data are reviewed in detail. Next, one challenging issue, classification in sparsely labeled networks, is reviewed extensively, and various solutions to this issue are discussed. Finally, the future development and expectation of networked data classification techniques are summarized. 2011 Vol. 24 (4): 527-537 [Abstract] ( 428 ) [HTML 1KB] [ PDF 918KB] ( 715 ) 538 A Fast Bi-Objective Non-Dominated Sorting Algorithm LIU Min, ZENG Wen-Hua, ZHAO Jian-Feng A fast bi-objective non-dominated sorting algorithm (BNSA) is proposed. An operator of forward comparison is designed to identify non-dominated individuals quickly. A sorting strategy according to need is proposed to avoid generating unnecessary non-dominated fronts. Then, the correctness of BNSA is proved and its time complexity is analyzed to be O(NlogN). Next, some comparable experiments are carried out on nine benchmark test problems for bi-objective optimization. Results of the experiments indicate that the proposed BNSA, for the most test problems, is faster than the other three non-dominated sorting algorithms. Furthermore, the BNSA, on all the test problems, has the best of accelerative effect, particularly when the number of evolutionary generations exceeds 400. In addition, the BNSA is concise and easy to be implemented. It can be incorporated into any multi-objective evolutionary algorithms based on non-dominated sorting to improve the running speed of bi-objective optimization. 2011 Vol. 24 (4): 538-547 [Abstract] ( 332 ) [HTML 1KB] [ PDF 641KB] ( 867 ) 548 Weight estimation for feature integration and saliency region extraction in modeling computation of visual selective attention Liu Qiong, Qin Shi-Yin According to the different importance of features during the feature integration process in modeling computation of visual selective attention, a method of weight estimation for different features is presented to highlight the acceptable saliency region of image in the bottom-up computational model. Firstly, the color, orientation and intensity features are extracted by mimicking the function of feature sensitive neurons of human primary visual cortex. Then the importance of each feature is estimated according to the generalized Gaussian distribution and the variance of its feature map. Finally, the saliency regions are extracted by weighted integration and normalization. The experimental results demonstrate that the proposed method outperforms traditional methods to meet the requirement of observers. 2011 Vol. 24 (4): 548-554 [Abstract] ( 379 ) [HTML 1KB] [ PDF 466KB] ( 926 ) 555 Study on Application of HMM to Online Signature Verification Based on Differences of Matched Segment ZOU Jie, WU Zhong-Cheng An approach of hidden markov model (HMM) to online signature verification is proposed, which uses difference values obtained by segmentation dynamic time wrapping (DTW) as observations of model. Firstly, the correspondences of the critical points in signatures are made by bidirectional backward-merging dynamic time wrapping algorithm. Then, the subtle differences are calculated by classical dynamic time wrapping algorithm. These differences are utilized to train the HMM. The meanings of models states are defined as degrees of similarity, and the HMM topology is ergodic. The validity of the proposed approach is verified on SVC2004 signatures database. 2011 Vol. 24 (4): 555-560 [Abstract] ( 496 ) [HTML 1KB] [ PDF 390KB] ( 753 ) 561 Query Expansion Based High Performance Chinese Voice Retrieval LI Wei, WU Ji, L Ping The aim of Chinese voice retrieval systems is to locate query texts in audio files fast and precisely. In a typical implementation of the system, voice files are recognized and stored in index. The system segments each query into a word sequence and uses the sequence to search. The mismatch between query segmentation and recognition can influence systems performance. To solve this problem, multiple segmentation results and prefix-suffix expansions have been used to broaden the original query. The retrieval process is on the basis of the expansions outputs. Query expansion generates a lot of outputs, which slows down the retrieval speed. In order to increase the systems efficiency, the Finite State Automata (FSA) is introduced to compress query expansions. And a Token-based search algorithm is used for fast search. Experimental results show that the query expansion leads the systems EER to improve about 50%~70% relatively. The FSA compresses the retrieval space, and raises the retrieval speed nearly 30 times. 2011 Vol. 24 (4): 561-566 [Abstract] ( 419 ) [HTML 1KB] [ PDF 378KB] ( 508 ) 567 Event Causal Relation Extraction Based on Cascaded Conditional Random Fields FU Jian-Feng, LIU Zong-Tian, LIU Wei, ZHOU Wen Traditional methods for event causal relation extraction covered only part of the explicit causal relation in the text. A method for event causal relation extraction is presented based on Cascaded Conditional Random Fields. The method casts the problem of event causal relation extraction as the labeling of event sequence. The Cascaded (Dual-layer) Conditional Random Fields is employed to label the causal relation of event sequence. The first layer of the Cascaded Conditional Random Fields model is used to label the semantic role of causal relation of the events, and then the output of the first layer is passed to the second layer for labeling the boundaries of the event causal relation. Experimental results show that this method not only covers each class of explicit event causal relation in the text, but also achieves good performance and the F-Measure of the overall performance arrives at 85.3%. 2011 Vol. 24 (4): 567-573 [Abstract] ( 735 ) [HTML 1KB] [ PDF 537KB] ( 987 ) 574 Time Series Classification Algorithm Based on Linear Segmentation and HMM YIN Rui, LI Xiong-Fei, LI Jun, PENG Hong The multi-segment linear (MSL) feature of the time series are collected, and a time series classification algorithm is proposed, which consists of derivative estimation function, linear segmentation method and DDHMM model (base on HMM). Firstly, the derivative estimation function and the linear segmentation method can be used together to detect the MSL feature. If they are matched, time series can be converted into observed sequence with a special structure. Next, the training observed sequences can be used to train DDHMM models. After training, the time series are classified through comparing the probability value of testing observed sequences generated by each model. The experimental results show that the proposed algorithm has a high accuracy when classifying the time series that match the MSL feature, and it has good performance in the classification on the UCI dataset and the actual projects. 2011 Vol. 24 (4): 574-581 [Abstract] ( 572 ) [HTML 1KB] [ PDF 580KB] ( 780 ) 582 Video Foreground Segmentation Based on Analysis of Spatial-Temporal Information Ming Hua-Qing, Chen Cong, Luo Rong-Hua, Zhu Jin-Hui An algorithm is proposed to segment foreground accurately from videos whose background is dynamic or whose foreground performs non-translational motion. Firstly, by regarding the change process of a single pixel as discrete-time signal, the video is segmented into foreground and background in a glancing way with temporal analysis using Gabor filter. Secondly, global color model and local color model are defined and built by clustering the color information of the background and foreground with mean-shift algorithm. Finally, a double-labeling method is used for fine segmentation of the foreground. Experimental results on several datasets prove that the proposed algorithm evidently improves the precision of the extracted foreground, especially in the cases that the background is dynamic or the foreground performs non-translational motion. 2011 Vol. 24 (4): 582-590 [Abstract] ( 432 ) [HTML 1KB] [ PDF 558KB] ( 682 ) 591 An Ensemble SVM Approach Integrated with Confidence for Detecting Bookmark Spam Zhang Fu-Zhi, Zhou Quan-Qiang The performance of existing methods for bookmark spam detection is decreased when there is less user profile information. An ensemble SVM approach integrated with confidence for detecting bookmark spam is proposed to solve this problem. The Bootstrap technology is firstly used to repeatedly sample the training data so as to get the subset of training samples for individual SVM. Then, sigmoid function is use to transform the standard output of SVM into a posterior probability which is used as the confidence of categories output. Finally, a method integrated with the confidence is proposed to aggregate the output of individual SVM, which is better than voting strategy. The experimental results show that the detection performance of the proposed approach outperforms the existing methods in the case of less user profile information. 2011 Vol. 24 (4): 591-600 [Abstract] ( 437 ) [HTML 1KB] [ PDF 402KB] ( 651 )

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