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.5, Published 2013-05-30 Orignal Article        Orignal Article 417 Qualitative Path Reasoning Based on Voronoi Diagram WANG Xiao-Dong,LIAO Shi-Zhong Qualitative motion reasoning is important in qualitative spatial reasoning research. Based on Voronoi diagram and its dynamic changes,a qualitative representation and reasoning method for motion paths is proposed. Firstly,the spatial relations among generators of Voronoi diagram are described by adjacent relations,the spatial positions and their conceptual neighborhood are defined,and the motion path is qualitatively represented by the sequence of conceptual-neighboring qualitative positions. Then,a qualitative path reasoning algorithm with an observer perspective is designed and implemented by the changes in edges of the dynamic Voronoi diagram and the proposed heuristic algorithm of shortest path between two qualitative positions. Finally,the experimental result and analysis demonstrate the validity of the proposed methods. 2013 Vol. 26 (5): 417-424 [Abstract] ( 766 ) [HTML 0KB] [ PDF 1353KB] ( 859 ) 425 Contracting Neighbor-Node-Set Approach for Solving Maximum Flow Problem in Directed Network ZHAO Shu,XU Xian-Sheng,HUA Bo,ZHANG Yan-Ping Maximum flow problem is widely applied in many fields. However,with the significant increase of network size,classic algorithms cannot solve maximum flow quickly and efficiently. In this paper,a method named Contracting Neighbor-node-set Approach (CNA) is presented to get its maximum flow approximately in a given directed flow network. Aiming at reducing the size of network,the method contracts some nodes and edges so that the classic algorithms can be used directly to approximately solve maximum flow problem with less time complexity. Firstly,the condition of contracting neighbor-node-set is given. Then,the algorithm is presented to construct the target network. Finally,the classic algorithms are applied on the target network to approximately get maximum flow of original network. The experimental results show that CNA not only obtains the maximum flow of original network with few errors,but also reduces the scale of the target flow network to half size of the original flow network averagely. 2013 Vol. 26 (5): 425-431 [Abstract] ( 504 ) [HTML 0KB] [ PDF 434KB] ( 598 ) 432 An Enhanced TranCo-Training Categorization Model with Transfer Learning TANG Huan-Ling,YU Li-Ping,LU Ming-Yu When unlabeled data draw from different distributions compared with labeled data in semi-supervise learning,the topic biases the target domain and the performance of semi-supervised classifier decreases. The transfer technique is applied to improve the performance of semi-supervised learning in this paper. An enhanced categorization model called TranCo-training is studied which combines transfer learning techniques with co-training methods. The transferability of each unlabeled instance is computed by an important component of TranCo-training according to the consistency with its labeled neighbors. At each iteration,unlabeled instances are transferred from auxiliary dataset according to their transfer ability. Theoretical analysis indicates that transfer ability of an unlabeled instance is inversely proportional to its training error,which minimizes the training error and avoids negative transfer. Thereby,the problem of topic bias in semi-supervised learning is solved. The experimental results show that TranCo-training algorithm achieves better performance than the RdCo-training algorithm when a few labeled data on target domain and abundant unlabeled data on auxiliary domain are provided. 2013 Vol. 26 (5): 432-439 [Abstract] ( 723 ) [HTML 0KB] [ PDF 619KB] ( 771 ) 440 Scene Classification Based on Global Optimized Framework JIN Tai-Song,LI Ling-Ling,LI Cui-Hua A scene classification algorithm based on global optimized framework is proposed. Firstly,the global scene feature named spatial envelop is obtained from the whole image,the visual word of each image block is extracted,and latent variable is defined to represent the semantic feature of the extracted visual word. Secondly,the structure graph of latent state is introduced to represent the context of visual words. In respect to scene classification strategy,objective function consisting of different potential functions is constructed in which potential functions are defined to measure the relevance of the variables including global scene feature,latent variables and scene category. Finally,the scene category of the image is determined when the global optimized solution of objective function is obtained. The experiments on the standard dataset demonstrate that the proposed algorithm achieves better results than the state-of-the-art algorithms. 2013 Vol. 26 (5): 440-446 [Abstract] ( 735 ) [HTML 0KB] [ PDF 638KB] ( 542 ) 447 A Matrix-Based Approach for Calculation of Knowledge Granulation WANG Lei,LI Tian-Rui The uncertainty is one of the hot issues in rough set theory. Knowledge granulation is a main approach to measure the uncertainty of knowledge systems. From the viewpoint of matrix,an approach for calculation of knowledge granulation,discernibility degree and attribute importance is studied,and the inherent meaning of matrix expression for the knowledge granulation is analyzed. Furthermore,the relations between the knowledge granulation and the equivalent relation matrix are revealed. Moreover,the hierarchical structure of knowledge granulation is proposed to discuss the variation of knowledge granulation under the addition or removal of a single attribute. Finally,combined with the update of the equivalent relation matrix under addition or removal of a signal attribute,the matrix-based computing method for attribute importance is applied in calculating the core set and the minimum attribute reduction. The numerical examples demonstrate the effectiveness of the proposed method on attribute reduction. 2013 Vol. 26 (5): 447-453 [Abstract] ( 699 ) [HTML 0KB] [ PDF 378KB] ( 606 ) 454 Depth Map Super-Resolution Based on the Local Structural Features of Color Image YANG Yu-Xiang,WANG Zeng-Fu It is convenient for time of flight camera to get the scene depth image,the resolution of depth image is very low due to limitations of the hardware,which can not meet the actual needs. In this paper,a method is proposed for solving depth map super-resolution problem. With a low resolution depth image as input,a high resolution depth map is recovered by using a registered and potentially high resolution camera image of the same scene. The depth map super-resolution problem is solved by developing an optimization framework. Specifically,the reconstruction constraint is applied to get the data term,and based on the fact that discontinuities in range and coloring tend to co-align,laplacian matrix and local structural features of high resolution camera images are used to construct the regularization term. The experimental results demonstrate that the proposed approach gets high resolution range image in terms of both its spatial resolution and depth precision. 2013 Vol. 26 (5): 454-459 [Abstract] ( 564 ) [HTML 0KB] [ PDF 1144KB] ( 817 ) 460 Design of 3D Latent-SVM and Application to Detection of Lesions in Chest CT WANG Qing-Zhu,KANG Wen-Wei,WANG Bin Accuracy of Computer Aided Detection (CAD) of lung lesions in chest CT may be affected by irregular shapes and simple texture of the lesions. To improve the poor performance of current CAD schemes,relative position from the suspected lesion to the whole lung area is added as a latent variable on the basis of traditional texture and shape features,which also participates in optimizing the SVM. Furthermore,considering 3D feature of the lung lesions,3D matrixes based SVM (3D SVM) is combined into the Latent SVM (L-SVM) to design 3D SVM with latent variables (3D-L-SVM). 150-case database from Jilin Tumor Hospital is used to validate the proposed algorithm. The performances of other three CAD schemes are compared on the same database. True positives of the 3D-L-SVM achieves 97.5% with the false positives of 9.21%. The experimental results verify the advantages of the proposed algorithm and effectiveness of assisting the radiologists. 2013 Vol. 26 (5): 460-466 [Abstract] ( 593 ) [HTML 0KB] [ PDF 821KB] ( 731 ) 467 A Clustering Algorithm for Transaction Sequences Based on Growth Patterns TANG Chun-Lei,DONG Jia-Qi,ZHU Bo-Ya,DAI Dong-Bo Mining and analysis of transaction sequences provide quantifiable schemes for decision makers to generate sales strategies. By studying the structure of transaction sequence sets according to the commodity sales amount and their variation trend,a kind of growth pattern is defined which reflects the variation trend of commodity price,as well as two methods of similarity measure,shifted window combined distance and angle vector distance,are defined. Based on those definitions,a clustering research is conducted by a goal function with time constraints. The experiments are conducted on the real commodity transaction sequence datasets. The results show that,combined with the growth patterns of two functions,it produces better clustering results under the condition of the time constraint,which could be well explained in practice. 2013 Vol. 26 (5): 467-473 [Abstract] ( 694 ) [HTML 0KB] [ PDF 453KB] ( 519 ) 474 Classification Algorithm of l2-norm LS-SVM via Coordinate Descent LIU Jian-Wei,FU Jie,WANG Shao-Lei,LUO Xiong-Lin The coordinate descent approach for l2 norm regulated least square support vector machine is studied. The datasets involved in the objective function for machine learning have larger data scale than the memory size has in image processing,human genome analysis,information retrieval,data management,and data mining. Recently,the coordinate descent method for large-scale linear SVM has good classification performance on large scale datasets. In this paper, the results of the work are extended to the least square support vector machine,and the coordinate descent approach for l2 norm regulated least square support vector machine is proposed. The vector optimization of the LS-SVM objective function is reduced to single variable optimization by the proposed algorithm. The experimental results on high-dimension small-sample datasets,middle-scale datasets and large-scale datasets demonstrate its effectiveness. Compared to the state-of-the-art LS-SVM classifiers,the proposed method can be a good candidate when data cannot fit in memory. 2013 Vol. 26 (5): 474-480 [Abstract] ( 391 ) [HTML 0KB] [ PDF 412KB] ( 502 ) 481 A Multi-Level Rough Set Model Based on Attribute Value Taxonomies YE Ming-Quan,HU Xue-Gang,HU Dong-Hui,WU Xin-Dong Most traditional studies on rough sets focus on finding attribute reduction and decision rules on the single level decision tables. However,attribute value taxonomies (AVTs) are usually predefined in applications and represented by hierarchy trees. Aiming at the attribute value taxonomies for condition attributes,the classical rough set model is extended to a multi-level rough set (MLRS) model combining with the full-subtree generalization mode. With decision table at different levels of generalization space,some properties of MLRS are obtained. Paralleling with attribute reduction based on positive region,a concept of attribute value generalization reduction in MLRS is introduced and the relations of generalization reduction and attribute reduction are analyzed. The computation of the generalization reduction in MLRS is proved to be a NP-hard problem. Then,a heuristic algorithm of generalization reduction based on the positive region of MLRS is proposed,which utilizes attribute value taxonomies to make top-down refinements. The attribute values are generalized to the optimal levels of their AVTs by the proposed algorithm,while the original positive region of the decision table keeps invariant. Theoretical analysis and simulation experiments illustrate that generalization reduction method improves the level and the generalization ability of knowledge discovery. 2013 Vol. 26 (5): 481-491 [Abstract] ( 672 ) [HTML 0KB] [ PDF 645KB] ( 602 ) 492 No-Reference Blurred Image Quality Assessment Based on Gray Level Co-occurrence Matrix SANG Qing-Bing,LI Chao-Feng,WU Xiao-Jun No-reference image quality assessment becomes a research hotspot recently. Based on gray level co-occurrence matrix,a no-reference blurred image quality assessment method is proposed which uses phase congruency feature learning. Firstly,this method generates phase congruency maps of testing images by Log Gabor wavelet. Secondly,it calculates the features of phase congruency map,which are entropy,energy,contrast,correlation and homogeneity based on gray level co-occurrence matrix. Finally,it predicts no-reference blurred image quality score by support vector regression (SVR) model training and learning. The experimental results on 4 public databases show that the predicted scores of the proposed method are in agreement with the subjective score,and it obtains a better evaluation index. 2013 Vol. 26 (5): 492-497 [Abstract] ( 444 ) [HTML 0KB] [ PDF 543KB] ( 1218 ) 498 Cell Segmentation in Microscopic Images of Mice BrainBased on Markov Random Field Theory SUN Li-Ye,HAN Jun-Wei,HU Xin-Tao,GUO Lei The neurons in sectioning microscope images of mice brain are important to biologists. Image segmentation algorithms are widely applied to automatically extract the neurons to facilitate further analysis. A method for cell segmentation in microscopic image of mice brain based on Markov Random Field (MRF) theory is proposed. Firstly,manually labeled images and original images are jointly analyzed to estimate the initial parameters in Gaussian Mixture Model,which significantly reduces the number of iterations and increases the precision of segmentation. Secondly,pixel intensity and distance between pixels are integrated into the conventional Potts model to improve the description of the quantitative relationship between pixels. The experimental results demonstrate that the proposed method improves the accuracy and the efficiency of cell segmentation compared to traditional methods. 2013 Vol. 26 (5): 498-503 [Abstract] ( 753 ) [HTML 0KB] [ PDF 907KB] ( 843 ) 504 Image Steganography Algorithm Based on Visual Attention and Local Complexity KANG Nian-Jin,CHEN Zhao-Jiong Visual attention is an important characteristic in the human perception system. However,most visual perception based steganography algorithms consider low-level factors only,such as brightness,contrast and masking effect. A gray image steganography algorithm is proposed based on visual attention and local complexity. The local complexity of the image is analyzed by standard deviation algorithm. Then,the Itti visual attention model is introduced into more complex areas,and the attention characteristics are quantitatively identified by visual entropy. Finally,the steganography is implemented by LSB method on different image blocks segmented through the hierarchies of visual attention and local complexity. The experimental results show that the proposed algorithm maintains good imperceptibility after embedding large-capacity information and resists the histogram contrast steganalysis. 2013 Vol. 26 (5): 504-512 [Abstract] ( 591 ) [HTML 0KB] [ PDF 0KB] ( 157 )

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