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....

2013 Vol.26 Issue.12, Published 2013-12-30 Orignal Article        Orignal Article 1089 Deep Belief Network Based Speaker Information Extraction Method CHEN Li-Ping, WANG Er-Yu, DAI Li-Rong, SONG Yan In i-vector based speaker verification system, it is necessary to extract the discriminative speaker information from i-vectors to further improve the performance of the system. Combined with the anchor model, a deep belief network based speaker-related information extraction method is proposed in this paper. By analyzing and modeling the complex variabilities contained in i-vectors layer-by-layer, the speaker-related information can be extracted with non-linear transformation. The experimental results on the core test of NIST SRE 2008 show the superiority of the proposed method. Compared with the linear discriminant analysis based system, the equal error rates(EER) of male and female trials can be reduced to 4.96% and 6.18% respectively. Furthermore, after the fusion of the proposed method with linear discriminant analysis, the EER can be reduced to 4.74% and 5.35%. 2013 Vol. 26 (12): 1089-1095 [Abstract] ( 515 ) [HTML 1KB] [ PDF 419KB] ( 986 ) 1096 Intent Waned Values and Attribute Reduction in Format Concept MA Yuan Attribute reduction is a hot topic in the area of concept lattice. The idea of intent waned values is used to prove that the set composed of arbitrary elements taken from intent waned values of each unique upper-neighbor concept is a reduction, thereby providing a simple algorithm of seeking all reductions, and this algorithm has a quite visual image in Hasse picture. Since the proposed algorithm only concerns unique upper-neighbor concepts, it can be finished by simply conducting some intersection calculation on context with no need to seek all the concept lattices. Therefore, the time and the space complexities of the proposed algorithm is still polynomial. 2013 Vol. 26 (12): 1096-1105 [Abstract] ( 357 ) [HTML 1KB] [ PDF 567KB] ( 596 ) 1106 An Evaluation Criterion of Salient Target Segmentation without Prior Knowledge in Infrared Image MIN Chao-Bo, ZHANG Jun-Ju, CHANG Ben-Kang, SUN Bin, LI Ying-Jie, LIU Lei A metric is presented for performance evaluation of salient target segmentation without prior knowledge in infrared image. According to the characteristics of salient targets in infrared images, three evaluation indexes, local illumination difference, degree of global salience and wave character of illumination, are proposed. They indicate the illumination contrast between the segmented foregrounds and corresponding local backgrounds, the degree of salience of the segmented foregrounds under the global backgrounds and the consistency of illumination of the segmented foregrounds, respectively. Finally, an evaluation criterion which consists of the proposed three indexes to measure the performance of salient target segmentation in infrared images is constructed comprehensively. Experimental results show that the proposed criterion provides an effective objective quality evaluation for salient target segmentation without prior knowledge in infrared images, which is consistent with human perception and objective fact. 2013 Vol. 26 (12): 1106-1114 [Abstract] ( 340 ) [HTML 1KB] [ PDF 1422KB] ( 550 ) 1115 Quotient Space Combination LANG Xian-Ji, WANG Jia-Yang The theory of quotient space is one of the three main methods of granular computing. The research of its combination is to find the relationship between the quotient space and the original space and to reduce the computational complexity by simplifying complex problems. The combination of domains is intended to implement granularity transform, fine or coarse, according to the different needs. The topological structure is the unique structure in the quotient space theory, and its combination has a variety of forms. In addition to the fine combination and the semi-order structure combination, the coarse combination and the converse quotient combination based on the converse-quotient topology are presented. The combination reflects the relationship between different topological structures. The combination of attribute functions lies in the formation of different equivalence relation of domain. These studies extend the basic theory of quotient space combination and make it better. 2013 Vol. 26 (12): 1115-1120 [Abstract] ( 313 ) [HTML 1KB] [ PDF 372KB] ( 494 ) 1121 Active Learning Based on Sparse Linear Reconstruction XIA Jian-Ming, YANG Jun-An, CHEN Gong The conventional active learning methods have one of the following defects:needing some labeled data selected randomly, ignoring the detail of the data structure, or requiring the fixed scale of the neighborhood to be set in advance. Therefore, a learning algorithm, active learning based on sparse linear reconstruction (SLR), is proposed based on the sparse representation model and the optimum experimental design method. Firstly, the sparse representation method is utilized to obtain the sparse reconstruction matrix. Then, the selection is realized with constraining the sparse reconstructive relationship among each data point and optimizing the reconstruction performance. Theory analysis and simulation results demonstrate that the proposed method selects the appropriate data points without any related prior information and does not need the fixed range between the nearby fields. Meanwhile, compared with the traditional methods such as neighborhood entropy, transductive experimental design and locally linear reconstruction, the proposed algorithm has better performance. 2013 Vol. 26 (12): 1121-1129 [Abstract] ( 304 ) [HTML 1KB] [ PDF 490KB] ( 739 ) 1130 A Survey of Hierarchical Classification Methods LU Yan-Ting, LU Jian-Feng, YANG Jing-Yu Hierarchical classification (HC), decomposing problem and organizing the classifiers according to the category hierarchy, is an efficient solution for multi-class classification problem. Depending on whether an explicit hierarchical relationship among categories is required, HC methods can be divided into two types. In this paper, the HC methods which do not require explicit hierarchical relationship among categories are reviewed systematically. Firstly, the basic framework of this type of methods is outlined. Then, the research progresses of several key techniques are elaborated and analyzed. Next, the related research work at home and abroad is described in detail from both algorithm and application perspectives. Finally, the existing methods are summarized and several future research directions are pointed out. 2013 Vol. 26 (12): 1130-1139 [Abstract] ( 760 ) [HTML 1KB] [ PDF 534KB] ( 5968 ) 1140 Motion Attention Fusion Model Based Video Target Detection and Extraction LIU Long, YUAN Xiang-Hui Aiming at the limitation of target detection and extraction algorithms under global motion scene, a target detection algorithm based on motion attention fusion model is proposed according to the motion attention mechanism. Firstly, the preprocess, such as accumulation and median filtering, is applied on the motion vector field. Then, according to the temporal and spatial characteristics of the motion vector, the motion attention fusion model is defined to detect moving target. Finally, the edge of the video moving target is extracted accurately by the morphologic operation and the edge tracking algorithm. The experimental results of different global motion video sequences show the proposed algorithm has better veracity and speedup than other algorithms. 2013 Vol. 26 (12): 1140-1145 [Abstract] ( 360 ) [HTML 1KB] [ PDF 1089KB] ( 576 ) 1146 Image Hierarchical Representation Model Based on LDA JIA Zhen-Hua, SIQING Ba-La The existing image hierarchical representation methods are strict in feed-forward style, and therefore it is not able to solve problems like local ambiguities well. In this paper, a probabilistic model is proposed to learn and deduce all layers of the hierarchy together. Specifically, a recursive probabilistic decomposition process is taken into account, and a generative model based on latent Dirichlet allocation with pyramidal multilayer structure is derived. Two important properties of the proposed probabilistic model are demonstrated: adding an additional representation layer to improve the performance of the flat model and adopting a full Bayesian approach which is better than a feed-forward implementation of the model. Experimental results on a standard recognition dataset show that the proposed method outperforms the existing hierarchical approaches, and it improves the classification and the learning accuracy with better performance. 2013 Vol. 26 (12): 1146-1153 [Abstract] ( 344 ) [HTML 1KB] [ PDF 569KB] ( 599 ) 1154 A Disparity Estimation Method Based on Curvature and Belief Propagation ZHAO Ge, LIN Lan, TANG Yan-Dong, WANG Yao-Nan The current widely used matching costs are sensitive to complex optical distortions, such as vignetting. By analyzing the camera model, the robustness of the curvature-based matching cost against the vignetting effect is proved. Then, the integral image is used to compute the curvature-based matching cost so that the speed of the algorithm is improved significantly. Finally, a regularization term is designed based on curvature constraint, and thus the over-smoothing of the disparity map is avoided by restricting the belief propagation in the depth discontinuity area. Experimental results verify the effectiveness of the proposed method. 2013 Vol. 26 (12): 1154-1160 [Abstract] ( 317 ) [HTML 1KB] [ PDF 667KB] ( 648 ) 1161 Hierarchical Clustering Based on a Bayesian Harmony Measure WEN Shun, ZHAO Jie-Yu, ZHU Shao-Jun Hierarchical clustering is an important data analysis technique. Traditional hierarchical clustering methods measure the similarity between two classes based on the Euclidean distance metric, and those methods can not deal with the overlapping between classes and the changes of the class density in range effectively. In this paper, a hierarchical clustering method based on a Bayesian harmony measure is presented. Instead of the Euclidean distance, the increase in the harmony degree is used to measure the similarity between two classes. The Bayesian harmony degree, introduced from the Bayesian Ying-Yang harmony learning theory, can measure the distribution of the entire dataset and guide the selection of the number of categories. The proposed method overcomes the drawbacks of the traditional methods. With the measure of Bayesian harmony degree, it becomes easier to select the threshold to terminate the merger of the hierarchical clustering and to generate the right number of categories. The experimental results on benchmark problems confirm the effectiveness of the proposed method. 2013 Vol. 26 (12): 1161-1168 [Abstract] ( 314 ) [HTML 1KB] [ PDF 1160KB] ( 909 ) 1169 Attribute Selection Method Based on Improved Discrete Glowworm Swarm Optimization and Fractal Dimension NI Zhi-Wei, XIAO Hong-Wang, WU Zhang-Jun, XUE Yong-Jian Attribute selection is an important method of data preprocessing in the field of data mining. An improved attribute selection method is proposed which combines discrete glowworm swarm optimization (DGSO) algorithm with fractal dimension. In this method,fractal dimension is taken as the evaluation criteria for attribute subsets and DGSO algorithm as a kind of search strategy. To analyze the feasibility and the effectiveness of the proposed method,six UCI datasets are used in the experiments,and the 10-fold cross validation and support vector machine algorithm are utilized to evaluate the classification accuracy before and after attribute selection. Then, different evaluation criteria and search strategies are compared and the parameters are analyzed in detail. The experimental results show that the proposed method has comparatively high feasibility and effectiveness. 2013 Vol. 26 (12): 1169-1178 [Abstract] ( 391 ) [HTML 1KB] [ PDF 533KB] ( 901 ) 1179 Geodesic Active Contour Based on LBF Method PAN Gai , GAO Li-Qun, ZHANG Ping It is difficult to get the real boundary of the object with poor boundary using the classical geodesic active contours. To deal with this problem, the local binary fitting(LBF) method and geodesic active contour are combined to propose a geodesic active contour based on LBF method in this paper. Firstly, the energy function of LBF method is normalized to replace the edge stopping function of geodesic active contour. Then, a gradient descent flow is established to urge an active contour to move at the boundary of the object. Finally, simulation experiments are implemented on 5 images with poor boundaries. Experimental results show the proposed model segment objects with poor edge correctly with anti-noise ability and it is insensitive to positions of the initial curve. The proposed model is superior to other improved geodesic active contours. 2013 Vol. 26 (12): 1179-1181 [Abstract] ( 368 ) [HTML 1KB] [ PDF 965KB] ( 665 )

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