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....

2015 Vol.28 Issue.4, Published 2015-04-30 Papers and Reports    Researches and Applications        Papers and Reports 289 Overlapping Community Discovery Based on Node Hierarchy and Label Propagation Gain CHEN Yu-Zhong, SHI Song, CHEN Guo-Long, YU Zhi-Yong The time complexity of multi-label propagation algorithm (MLPA) is nearly linear. However, when it is applied to overlapping community discovery, the accuracy and the stability of MLPA are poor. Inspired by the idea that overlapping nodes are more probable to appear in the boundary regions of different communities, an overlapping community discovery algorithm based on node hierarchy and label propagation gain is proposed in this paper. Firstly, the improved single label propagation with node centrality and community distribution constraints is utilized to unfold preliminary non-overlapping communities and centrality values of nodes are calculated by local information in the propagation process simultaneously. Furthermore, node hierarchy partition function is defined according to centrality values of nodes and employed to mark the hierarchy of each node in its respective community. Finally, based on the label propagation gain among nodes, a new multi-label updating rule is designed to obtain the final overlapping communities. Extensive experimental results on synthetic and real-world networks validate that the proposed algorithm effectively improves the accuracy and stability. 2015 Vol. 28 (4): 289-298 [Abstract] ( 634 ) [HTML 1KB] [ PDF 559KB] ( 801 ) 299 Recurrent Neural Network Language Model Based on Word Vector Features ZHANG Jian, QU Dan, LI Zhen The recurrent neural network language model(RNNLM) solves the problems of data sparseness and dimensionality disaster in traditional N-gram models. However, the original RNNLM is still lack of long dependence due to the vanishing gradient problem. In this paper, an improved method based on contextual word vectors is proposed for RNNLM. To improve the structure of models, a feature layer is added into the input layer. Contextual word vectors are added into the model with feature layer to reinforce the ability of learning long-distance information during the training. Experimental results show that the proposed method effectively improves the performance of RNNLM. 2015 Vol. 28 (4): 299-305 [Abstract] ( 744 ) [HTML 1KB] [ PDF 416KB] ( 2473 ) 306 Intelligent Optimization Strategy for Virtual Machine Placement in Data Center NI Zhi-Wei, LIANG Ting, WU Zhang-Jun, XIAO Hong-Wang Virtual machine placement problem is one of the key issues affecting the performance of data centers. An intelligent optimization strategy for virtual machine placement in the data center is proposed with comprehensive consideration of resource wastage, power consumption and load balance. Firstly, a multi-objective mathematical model for the virtual machine placement optimization is built by the strategy. Secondly, the placement problem is abstracted as the bin packing problem. Finally, an optimization strategy based on the improved adaptive discrete glowworm swarm optimization is put forward. Simulation experiment results show that the proposed adaptive discrete glowworm swarm optimization has good robustness and convergence rate, and the proposed intelligent optimization strategy solves the virtual machine placement problem effectively. 2015 Vol. 28 (4): 306-315 [Abstract] ( 546 ) [HTML 1KB] [ PDF 715KB] ( 1081 ) 316 Map Building Method Based on Hierarchical Temporal Memory ZHANG Xin-Zheng, MAI Xiao-Chun, ZHANG Jian-Fen A map building method based on hierarchical temporal memory (HTM)is proposed. The mapping problem is treated as scene recognition. The map is composed of a series of scenes being the outputs of HTM network. Firstly, the position invariant robust feature (PIRF) is extracted from the obtained images and then the PIRFs are applied to build the visual vocabulary. Secondly, according to the visual vocabulary PIRF descriptors of an image are projected to the vector of visual word occurrences. Multiple visual word occurrences vectors are formed as a sequence of visual word occurrences. This sequence is inputted to HTM to implement the environment map learning and building and closed loop scenes recognition. The performance of the proposed mapping method is evaluated by two experiments. The results show that the proposed strategy based on HTM is effective for map building and closed loop detection. 2015 Vol. 28 (4): 316-326 [Abstract] ( 574 ) [HTML 1KB] [ PDF 1452KB] ( 712 ) 327 An Attribute Reduction Algorithm Based on Granular Computing and Discernibility JI Su-Qin, SHI Hong-Bo, Lü Ya-Li In traditional attribute reduction algorithms, all the data are loaded into the main memory once, which is hard to adapt to the big data analyses. Aiming at this problem, an attribute reduction algorithm based on granular computing and discernibility is proposed. An original large-scale datset is divided into small granularities by applying stratified sampling in statistics, and then attributes are reduced on each small granularity based on discernibility of attribute. Finally, all the reductions on small granularities are fused by weighting. Experimental results show that the proposed algorithm is feasible and efficient for attribute reduction on massive datasets. 2015 Vol. 28 (4): 327-334 [Abstract] ( 476 ) [HTML 1KB] [ PDF 470KB] ( 652 ) 335 Cooperative Multi-robot Observation of Multiple Moving Targets Based on Contribution Model YANG Jian-Hua, ZENG Wen-Jia, WU Zhao-Hui How to reduce the overlap observation phenomena and improve the average observation rate at the same time is a complicated problem of cooperative multi-robot observation of multiple moving targets. An approach based on contribution for cooperative multi-robot observation of multiple moving targets (C-CMOMMT) is proposed. Each robot is endowed by the C-CMOMMT algorithm with a contribution value derived from the number of assigned targets to it. Robots with low contribution receive strengthened repulsive forces from all others. Besides, the operating distances of all repulsive forces are expanded, and robots with high contribution receive weakened attractive forces from low-weighted targets. With these three methods the overlap observation phenomena are reduced. To decrease the target loss, robots with high contribution receive feeble repulsive forces, and thus the side effects become weak. Consequently, the robots are decentralized and the overlap observing phenomena are dwindled. The average observation rate, the standard deviation and entropy of the positions of mobile robots are introduced to systematically evaluate the performance and the degree of overlap observing. Results show that C-CMOMMT improves the average observation rate and dwindles the overlap observing phenomena and it is more effective than A-CMOMMT and B-CMOMMT. 2015 Vol. 28 (4): 335-343 [Abstract] ( 489 ) [HTML 1KB] [ PDF 599KB] ( 593 ) Researches and Applications 344 Superpixel Graph Cuts Rapid Algorithm for Extracting Object Contour Shapes ZHANG Rong-Guo, LIU Xiao-Jun, DONG Lei, LI Fu-Ping, LIU Kun A rapid algorithm based on level set framework is presented for extracting object contour shapes. Firstly, initial seeds are placed in an image plane evenly. Through setting superpixel evolution forces, superpixels with similar region features are generated. The image segmented by these superpixels maintains geometric characteristics of object contour shapes and in the meantime prevents overlap between superpixel regions. Secondly, based on the relationship of superpixel labeling and Heaviside function, optimization model of the Mumford-Shah energy function is built by using graph cuts. Finally, geometric shapes of the object contour can be extracted by superpixel graph cuts. Experimental results show that the number of superpixels is reduced greatly, converted optimization model satisfies requirements of graph cuts against energy function optimization, and min-cut/max-flow method does not need to solve differential equations. Higher extracting effectiveness of object contour shapes and extracting efficiency are ensured by all these measures. 2015 Vol. 28 (4): 344-353 [Abstract] ( 559 ) [HTML 1KB] [ PDF 1653KB] ( 826 ) 354 Spherical Cover Classification Algorithm Based on Manifold Dimension Reduction Space of Local and Global Mapping HU Zheng-Ping, DU Li-Cui, ZHAO Shu-Huan To explore the intrinsic structure and the low dimensional representation of high dimensional data and find explicit mapping in some manifold learning algorithms, spherical cover classification algorithm based on manifold dimension reduction space of local and global mapping is proposed. The mapping model combining local information and global information is extracted firstly. The local laplacian matrix and the global laplacian matrix are optimized separately. The low dimensional representation of training data is obtained by eigen-decomposition of the laplacian matrix. Then the low dimensional representation of testing data is obtained by kernel mapping. Finally, the spherical cover classification model in low dimensional space is constructed. Extensive experiments are conducted on MNIST dataset, YaleB face dataset and AR dataset, and the results verify the effectiveness of the proposed algorithm and its value in the application fields. 2015 Vol. 28 (4): 354-360 [Abstract] ( 458 ) [HTML 1KB] [ PDF 609KB] ( 608 ) 361 Real-Time Object Tracking Algorithm Based on Adaptive Compressive Feature Selection QI Mei-Bin, LU Lei, YANG Xun, YANG Yan-Fang, JIANG Jian-Guo Low dimensional features adopted by compressive tracking algorithm can not reconstruct the object effectively. To solve this problem, a real-time object tracking algorithm based on adaptive compressive feature selection is proposed in this paper. The high dimensional features meeting the requirement of object reconstruction are extracted. Then the lower dimensional features with a higher discrimination are selected as appearance model of the object to reduce the computational complexity. To select features adaptively a difference method is adopted to control the feature dimensionality. The experimental results demonstrate that the proposed algorithm are more robust and effective in real time than other state-of-the-art tracking algorithms. 2015 Vol. 28 (4): 361-368 [Abstract] ( 504 ) [HTML 1KB] [ PDF 1904KB] ( 695 ) 369 An Emotion Decision-Making Model in Multi-agent Environment LIN Jun-Huan, LIU Zhen, CHEN Yue-Fen An emotion decision-making model consisting of cognition layer and emotion layer is constructed, the cognition layer is implemented in the Nash-Q algorithm, and the emotion layer is based on the theory of emotion memory and evaluation. The emotion space includes happiness, sadness, fear, boredom. The stimulus-to-emotion mapping model, emotion-to-action mapping model and the evaluation model of action credibility for each emotion are built respectively. The proposed model is applied to two-agent grid decision-making experiment. The results show that the convergence speed is higher when the Nash-Q algorithm is combined with emotional layer, and the model can effectively avoid local optimum. The model keeps better balance between conservation and searching in online learning. 2015 Vol. 28 (4): 369-376 [Abstract] ( 541 ) [HTML 1KB] [ PDF 562KB] ( 576 ) 377 Event Boundary Detection Method Based on Wireless Sensor Network and Linear Neural Network WU Peng-Fei, LI Guang-Hui, ZHU Hong, ZENG Song-Wei, LU Wen-Wei Environmental monitoring is a typical application in wireless sensor network (WSN), and event boundary detection is important for environmental monitoring. In this paper, a temporal-spatial data model of WSN is established, and then an event boundary detection method based on the linear neural network is presented. Firstly, the temporal correlation of data stream is analyzed, and the abnormal data set is determined based on linear neural network technique. Then, the event boundary is detected by using the spatial correlation of data stream between the neighbor nodes, and both the fault nodes and the event boundary nodes can be found. Thus, the location and the size of the event region can be estimated. Theoretical analysis and experimental results show that the proposed method has a high accuracy of fault node and event boundary detection and a low false positive rate. 2015 Vol. 28 (4): 377-384 [Abstract] ( 520 ) [HTML 1KB] [ PDF 814KB] ( 932 )

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