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....

2017 Vol.30 Issue.3, Published 2017-03-31 Orignal Article        Orignal Article 193 Automatic Image Annotation Combining Semantic Neighbors and Deep Features KE Xiao, ZHOU Mingke, NIU Yuzhen In the traditional image annotation methods, the manual selection of features is time-consuming and laborious. In the traditional label propagation algorithm, semantic neighbors are ignored. Consequently visual similarity and semantic dissimilarity are caused and annotation results are affected. To solve these problems, an automatic image annotation method combining semantic neighbors and deep features is proposed. Firstly, a unified and adaptive depth feature extraction framework is constructed based on deep convolutional neural network. Then, the training set is divided into semantic groups and the neighborhood image sets of the unannotated images are set up. Finally, according to the visual distance, the contribution value of each label of the neighborhood images is calculated and the keywords are obtained by sorting their contribution values. Experiments on benchmark datasets show that compared with the traditional synthetic features, the proposed deep feature possesses lower dimension and better effect. The problem of visual similarity and semantic dissimilarity in visual nearest neighbor annotation method is improved, and the algorithm effectively enhances the accuracy and the number of accurate predicted tags. 2017 Vol. 30 (3): 193-203 [Abstract] ( 905 ) [HTML 1KB] [ PDF 1344KB] ( 611 ) 204 Discrete Multi-objective Quantum Particle Swarm Optimization Clustering Algorithm ZHAGN Yong, WANG Qing, XIA Changhong, SUN Xiaoyan, GONG Dunwei Clustering is a significant data processing technique in data mining field. An improved multi-objective clustering algorithm based on quantum particle swarm optimization is proposed. Firstly, an integer coding strategy is introduced for the unknown class centers. Then, an effective particle swarm optimization strategy is designed based on Canopy strategy to predict the number of class centers. An improved discrete quantum update formula is defined to update the particle position by introducing versus, and and difference operators. Finally, the proposed algorithm is applied to seven real datasets and compared with two typical single-objective clustering algorithms and three multi-objective clustering algorithms. Experimental results demonstrate the effectiveness of the proposed algorithm. 2017 Vol. 30 (3): 204-213 [Abstract] ( 661 ) [HTML 1KB] [ PDF 961KB] ( 399 ) 214 Fuzzy Clustering Image Segmentation Based on Neighborhood Constrained Gaussian Mixture Model ZHAO Quanhua, ZHANG Hongyun, ZHAO Xuemei, LI Yu The characteristics of data can not be simulated in the traditional fuzzy clustering method effectively. Gaussian mixture model with neighbor constraints is introduced to solve the problem. Gaussian distribution is used to characterize the statistical characteristics of spectral measure. The correlation between the pixels and their neighborhood pixels are defined as prior probability and used as weight coefficients of each component in Gaussian mixture model. Finally, a Gaussian mixture model with neighborhood constraints in feature field is constructed. Log weighted Gaussian component in the mixture model is used as dissimilar measurement between the pixels and clusters, and a fuzzy clustering objective function is constructed based on Gaussian mixture model. Neighborhood constraints are introduced as a weight of component in traditional Gaussian mixture model and combined with fuzzy clustering method. Thus, the problem of multi-peak distribution of data is solved. Finally, the accuracy of the proposed algorithm is verified by experiments. 2017 Vol. 30 (3): 214-224 [Abstract] ( 611 ) [HTML 1KB] [ PDF 1036KB] ( 503 ) 224 Multi-objective Particle Swarm Optimization Algorithm Based on Balance Search Strategy GENG Huantong, CHEN Zhengpeng, CHEN Zhe, ZHOU Lifa Considerating the importance of balancing global and local search for multi-objective particle swarm optimization algorithm(MOPSO) to obtain the complete and uniform Pareto front(PF), a balance search strategy is designed and an improved multi-objective particle swarm optimization algorithm (bsMOPSO) is proposed.The strategy is composed of two novel search sub-strategies. In the local search sub-strategy, self-exploitation of archive set is designed to achieve local search involving the entire Pareto front by disturbing fixed ratio of uniform particles in archive set with Cauchy mutation. In the global search sub-strategy, guided search by the best boundary particle is designed through using the optimal boundary particle as the global optimal solution, and therefore more boundary areas of each objective function are searched by part of particle swarm. By comparing five algorithms on the series of ZDT and DTLZ test functions, the results demonstrate that bsMOPSO achieves better Pareto optimal convergence and distribution. 2017 Vol. 30 (3): 224-234 [Abstract] ( 520 ) [HTML 1KB] [ PDF 3543KB] ( 736 ) 235 Group Sample Learning to Rank Approach Based on Likelihood Loss Function LIN Yuan, XU Bo, SUN Xiaoling, LIN Hongfei, XU Kan Group sample used for training the ranking model provides a new idea to construct learning to rank methods. In this paper, the new loss function is constructed for group samples to train the learning to rank model. The preference-weighted loss function and the initial ranking list optimization are employed to construct a new group learning to rank method based on neural network. Experimental results show that the proposed approach is effective in improving ranking performance. 2017 Vol. 30 (3): 235-241 [Abstract] ( 484 ) [HTML 1KB] [ PDF 661KB] ( 507 ) 242 Twice Learning Based Semi-supervised Dictionary Learning for Software Defect Prediction ZHANG Zhiwu, JING Xiaoyuan, WU Fei When the previous defect labels of modules in software history warehouse are limited, building an effective prediction model becomes a challenging problem. Aiming at this problem, a twice learning based semi-supervised learning algorithm for software defect prediction is proposed. In the first stage of learning, a large number of unlabeled samples are labeled with probability soft labels and extended to the labeled training dataset by using sparse representation classifier. Then, on this dataset discriminative dictionary learning is used for the second stage of learning. Finally, defect proneness prediction is conducted on the obtained dictionary. Experiments on the widely used NASA MDP and PROMISE AR datasets indicate the superiority of the proposed algorithm. 2017 Vol. 30 (3): 242-250 [Abstract] ( 530 ) [HTML 1KB] [ PDF 712KB] ( 430 ) 251 Subspace Clustering Based on Collaborative Representation FU Wenjin, WU Xiaojun, DONG Wenhua, YIN Hefeng The coefficient matrix solved by sparse subspace clustering(SSC) is too sparse and the coefficient matrix solved by least squares regression for subspace clustering(LSR) is too dense. Aiming at these problems, subspace clustering based on collaborative representation(SCCR) is proposed. The advantages of SSC and LSR are combined. The l1 norm and the Frobenius norm are introduced into an objective function. The coefficient solved by SCCR can group the correlated data within cluster like LSR and eliminate the connection between clusters like SSC. Then, the affinity matrix constructed by this coefficient matrix is applied to spectral clustering. The experimental results demonstrate that SCCR improves the performance of clustering. 2017 Vol. 30 (3): 251-259 [Abstract] ( 484 ) [HTML 1KB] [ PDF 896KB] ( 479 ) 260 Method of Improving Practicability of Indoor Visual Odometry PENG Tianbo, WANG Hengsheng, ZENG Bin Aiming at the controversy of the real-time performance, robustness and accuracy in visual odometry, method of improving practicability of indoor visual odometry is put forward to tackle the problem. The corner features of every image in the sequence are obtained using graphics processing unit based oriented FAST and rotated BRIE algorithm and matched using K Nearest neighbor algorithm to reduce the computation time. According to the measurement range of Kinect, points with high measurement error are rejected. To solve the movement of the camera between two frames, the estimation of movement parameters are firstly obtained with efficient perspective-n-point algorithm. Then, they are used as the initial value of Levenberg-Marquedt algorithm to refine the parameters. Random sample consensus is used to reject outliers during the computation of the camera movement. The experimental results show that the proposed method is effective for the accuracy improvement of the motion trajectory calculation. 2017 Vol. 30 (3): 260-268 [Abstract] ( 475 ) [HTML 1KB] [ PDF 996KB] ( 556 ) 269 Person Re identification Based on Multi feature Fusion YUAN Li,TIAN Ziru Due to variations in pose and illumination condition, the appearance of a person can be significantly different in two views and therefore the performance of person re identification is degraded. In this paper, a feature fusion method for person re identification is proposed including HSV color feature， color histogram feature and texture feature extracted by the histogram of oriented gradient descriptor. The specific process is divided into the training phase and the recognition phase. In the training phase, the feature descriptors of each image in the reference dataset are firstly extracted, and then a correlation matrix of the image features from two cameras is learned using canonical correlation analysis. As for re identification, the feature descriptors of each image in the gallery dataset and the probe dataset are firstly extracted, and then they are transformed by the correlation matrix. Finally， re identification is implemented by measuring the similarity between the gallery image descriptor and the probe image descriptor. Experimental results on three datasets show that the proposed method outperforms the state of art approaches. 2017 Vol. 30 (3): 269-278 [Abstract] ( 754 ) [HTML 1KB] [ PDF 1694KB] ( 546 ) 279 Preference Feature Extraction Based on Column Union Row Matrix Decomposition LEI Hengxin, LIU Jinglei Preference features can not be accurately analyzed and explained by singular value decomposition. Aiming at these problems, a column union row(CUR) matrix decomposition method is proposed to acquire a low-rank approximation of the original matrix M (user preferences for products) and extract the potential preferences of users and products. The statistics leverage score of matrix M is calculated firstly. And then, several rows and columns with higher scores are extracted to constitute low-dimensional matrix C and matrix R. Subsequently, the matrix U is constructed approximatively according to matrix M, C and R. By the proposed method, the extraction problem of preference feature in a high-dimensional space is transformed to the matrix analysis problem in a lower dimensional space. As a consequence, the CUR decomposition has better accuracy and interpretability. Finally, the theoretical analysis and experiment indicate that compared with the traditional decomposition methods, the CUR matrix decomposition method has higher accuracy, better interpretability and higher compression ratio for extracting preference feature. 2017 Vol. 30 (3): 279-288 [Abstract] ( 576 ) [HTML 1KB] [ PDF 1055KB] ( 961 )

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