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....

2016 Vol.29 Issue.6, Published 2016-06-30 Papers and Reports    Researches and Applications        Papers and Reports 481 Flexible Calibration Method with High Accuracy for Dynamic Focusing ZHOU Jiali, JIA Lushuai, WU Min The camera intrinsic parameters keep changing during the dynamic focusing process and different focusing states. By linking an angle sensor on the focusing ring, a camera model using the increment of the principal distance as the parameter is proposed based on the dynamic focusing process. The camera intrinsic parameters under arbitrary focusing state and object distance can be solved. A highly robust method is put forward by the customized stereo calibration target with three planes. Compared with the traditional model, no additional constraint condition is required and the proposed model has a strong operability. Experiments show the stronger image rectification ability of the proposed model and the corresponding calibration method. Furthermore, the precision of photogrammetry is improved by the proposed method. 2016 Vol. 29 (6): 481-491 [Abstract] ( 453 ) [HTML 1KB] [ PDF 2888KB] ( 618 ) 492 Clustering Segmentation of Dominant Colors Based on Salient Feature Fusion SIMA Haifeng, MI Aizhong, WANG Zhiheng, DU Shouheng Aiming at the fault segmentation caused by color density clustering segmentation model, a dominant colors clustering image segmentation algorithm is proposed based on visual saliency. Firstly, according to the spatial color information and Mean-shift smoothing results, the global saliency and region saliency of the image are computed and fused as the constraints of spatial clustering. Then, kernel density estimation is employed to compute dominant colors of image as initial clusters and the salient features are taken as regulated factors for clustering segmentation. Finally, regions are merged for final segmentation. The experiments are implemented on the standard segmentation database and the proposed algorithm is compared with several algorithms. The experimental results show the higher precision of the proposed algorithm on region contours. The proposed algorithm makes good use of the salient feature of image, reduces the inconsistency of the clustering results, and improves the accuracy of pixel clustering and the robustness of the segmentation. 2016 Vol. 29 (6): 492-503 [Abstract] ( 501 ) [HTML 1KB] [ PDF 3328KB] ( 1129 ) 504 Combinatorial Semi-supervised Incremental Support Vector Machine Learning Algorithm GUO Husheng, WANG Wenjian, PAN Shichao Incremental support vector machine (ISVM) has difficulty in selecting the best incremental sample during each incremental learning step, and therefore the generalization performance of the model is weak. To solve this problem, combinatorial semi-supervised incremental support vector machine learning algorithm (ICS3VM) is proposed. The best incremental sample is selected by combinatorial labeling of the large scale unlabeled samples in batches. The most valuable unlabeled samples in the classification margin are added into the training set each time to correct the model. Meanwhile, the label with the largest margin is regarded as the final label to ensure the accuracy. The experiment on the standard datasets shows the good generalization performance and the high learning efficiency of the proposed ICS3VM. 2016 Vol. 29 (6): 504-510 [Abstract] ( 381 ) [HTML 1KB] [ PDF 426KB] ( 505 ) 511 Person Re-identification Based on Regularization of Independent Measure Matrix QI Meibin, WANG Yunxia, TAN Shengshun, LIU Hao, JIANG Jianguo To solve the over-fitting problem caused by less training samples in the current person re-identification method based on distance metric learning, a person re-identification algorithm based on regularization of independent measure matrix is proposed. Firstly, the features extracted from four different color spaces are used to learn four different measure matrices. Then, the corresponding matrixes are regularized respectively, and the similarity of testing examples is measured by the regularized matrices. Finally, the final similarity is obtained by fusing results of the similarity measure. Experimental results show the improvement of the proposed method in performance for the over-fitting problem caused by less training samples. 2016 Vol. 29 (6): 511-518 [Abstract] ( 460 ) [HTML 1KB] [ PDF 427KB] ( 598 ) 519 Improved Biclustering Algorithm Based on Weighted Mean Square Residual LIU Wenhua, LIANG Yongquan, FENG Zheng Existing biclustering algorithms can hardly discover biclusters with overlapping structures. Consequently, the correct bicluster structures hidden in gene expression data can not be effectively found. Moreover, the influence of the importance of the different conditions on the bicustering result is not taken into account in the process of adding and deleting conditions. An improved biclustering algorithm based on weighted mean square residual(IBWMSR) is proposed to overcome the above defects. The gene sets are firstly partitioned into initial biclusters by using fuzzy partition and the fuzzy partition is controlled by overlapping ratio and the membership of the genes. Then, the weights of the conditions in each bicluster are iteratively updated in the process of minimizing the objective function. Finally, the bicluster set is obtained after adding the genes satisfying the constraints and removing the genes producing inconsistency fluctuation. Theexperiment shows that the proposed algorithm generates the biclusters with similar expression level of different sizes and restricts the overlapping ratio to a reasonable range. 2016 Vol. 29 (6): 519-526 [Abstract] ( 454 ) [HTML 1KB] [ PDF 520KB] ( 599 ) Researches and Applications 527 Algorithm of Maximum Correntropy Based on l2-Regularization in Individual Communication Transmitter Identification TANG Zhe, LEI Yingke To measure the similarity between the fine features of communication transmitters, an algorithm of individual communication transmitter identification using maximum correntropy based on l2-regularization is put forward. Firstly, the square integral bispectra of the communication transmitter signal is extracted to characterize the individual differences of communication transmitters, and then optimization model using maximum correntropy criterion based on l2-regularization is constructed. Next, the half-quadratic technique is used to transform the nonlinear optimization problem into a weighted linear least squares problem. Finally, the sparse coefficient can be computed by active set algorithm, and then the classifier is constructed by mining the discriminative information of coefficient for the communication transmitters identification. The feasibility and the effectiveness of the proposed algorithm are verified on the real datasets collected from the FM radios with same manufacturer and model. 2016 Vol. 29 (6): 527-533 [Abstract] ( 314 ) [HTML 1KB] [ PDF 480KB] ( 460 ) 534 Clothing Recognition Based on Clothing Co-occurrence Information and Multi-task Learning HAO Zhifeng, LING Suiyi, WEN Wen, CAI Ruichu, YUAN Chang Multi-task learning (MTL) ignores the influence of prior probability on the process of learning. Aiming at this problem, an approach for clothing recognition based on MTL and the co-occurrence information of the clothes categories (CA-MTL) is proposed. The prior constraint term is incorporated into the MTL model to integrate the co-occurrence information of the clothes categories. The original extended gradient method is modified correspondingly and the performance of the clothing classifiers is thus enhanced. Experimental results show that the average performance of CA-MTL outperforms those of single task learning, neural network and traditional multi-task learning. Furthermore, the training results of the proposed model are convenient for visualization and can be used for feature selection. 2016 Vol. 29 (6): 534-541 [Abstract] ( 391 ) [HTML 1KB] [ PDF 518KB] ( 876 ) 542 Semi-supervised Acoustic Modeling Based on Perplexity Data Selection XIE Chuandong, GUO Wu For acoustic modeling of small languages with rare resource, a perplexity based approach is proposed to select unsupervised data in the decoding transcription and retrain the acoustic model. The large unsupervised corpus is decoded using the initial acoustic model trained with a small amount of labeled data, and the perplexity between the decoded text and the training set is calculated. Then, the selected data similar to the labeled data are used to train the acoustic model along with the labeled data. To improve the correctness of the decoded unsupervised data,the final network parameters of acoustic model are adjusted by only using the correct labeled data in the last iteration during the training of model parameters based on deep neural network. In the VLLP recognition task of Swahili provided by NIST 2015 open keyword search competition, the proposed approach can improve the recognition rate compared with other methods. 2016 Vol. 29 (6): 542-547 [Abstract] ( 344 ) [HTML 1KB] [ PDF 367KB] ( 571 ) 548 Functional Module Detection Based on Multi-label Propagation Mechanism in Protein-Protein Interaction Networks HAN Yue, JI Junzhong, YANG Cuicui Due to the fast and efficient solution of multi-label propagation algorithm in detecting community for social network, a functional module detection based on multi-label propagation mechanism in protein-protein interaction (PPI) networks (MLP-FMD) is proposed by merging multi-source protein biological knowledge. Firstly, the labels of nodes are initialized by using the functional and structural information of a PPI Network. Then, the co-expression of the protein is calculated by using the gene expression data and thus the label set of the nodes is constructed, and the label is selected to achieve the true and reliable transmission among the nodes. Finally, the nodes with same identifier are divided into the same functional module, and the final result is obtained. Experiments show good time performance and a certain competitive ability of the detection accuracy of the proposed algorithm. 2016 Vol. 29 (6): 548-557 [Abstract] ( 441 ) [HTML 1KB] [ PDF 567KB] ( 473 ) 558 Soft Subspace Clustering Algorithm Based on Quantum-Behaved Particle Swarm Optimization XU Yajun, WU Xiaojun Soft subspace clustering algorithm frequently falls into local optimum during searching clustering center point. A fuzzy clustering algorithm is proposed based on the framework of soft subspace clustering, and it integrates quantum-behaved particle swarm optimization (QPSO) algorithm into gradient descent method to optimize the objective function in soft subspace clustering. By the characteristic of searching global optimum in the QPSO algorithm, global optimal center points are solved in the subspace, and then by the high convergence speed of the gradient descent method, fuzzy weights and membership degree matrices of sample points can be obtained. Finally, the optimal clustering results of sample points are obtained. Experiment is carried out on UCI dataset and the results demonstrate the improvement in accuracy as well as the stability of the clustering results of the proposed method. 2016 Vol. 29 (6): 558-566 [Abstract] ( 398 ) [HTML 1KB] [ PDF 449KB] ( 490 ) 567 Classifier Chain Algorithm Based on Multi-label Importance Rank LI Na, PAN Zhisong, ZHOU Xingyu The learning performance of the classifier chain algorithm often decreases due to the random prediction order of multiple labels in the classifier chains. Moreover, the error information is disseminated. With the consideration of the order of labels in a chain, a classifier chain algorithm based on multi-label importance rank is proposed. The information of interaction between the markers is used as a prerequisite to measure the label importance. On the basis of correlation, the labels are sorted according to their importance, and the ranking results are regarded as the classifier order in classifier chain algorithm. Thus, the problem of multi-label prediction sequence is solved. Experiments show that the proposed algorithm is more stable and efficient for multi-label classification compared with some state-of-the-art methods. 2016 Vol. 29 (6): 567-575 [Abstract] ( 520 ) [HTML 1KB] [ PDF 483KB] ( 945 )

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