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.9, Published 2016-09-30 Papers and Reports    Researches and Applications        Papers and Reports 769 Comparative Study of Light Field Depth Estimation GAO Jun, WANG Lijuan, ZHANG Xudong, ZHANG Jun To achieve accurate depth estimation by using light field data effectively, the light field depth estimation is reviewed in this paper. Firstly, the basic theory of light field is expounded, and the light field depth estimation methods are classified into three categories, methods based on epipolar plane image, multiview and refocusing. Next, the effects of illuminant variations on the performance of depth estimation are compared on synthetic datasets. Besides, a more comprehensive and challenging light field dataset is constructed, and the effect of complex scenes on the performance of depth estimation is qualitatively and quantitatively analyzed on the light field benchmark dataset and LytroDataset. Furthermore, the development of this field is pointed out. 2016 Vol. 29 (9): 769-779 [Abstract] ( 892 ) [HTML 1KB] [ PDF 1839KB] ( 1405 ) 780 Unstable Cut-Points Based Sample Selection for Large Data Classification WANG Xizhao, XING Sheng, ZHAO Shixin When the traditional sample selection methods are used to compress the large data, the computational complexity and large time consumption are high. Aiming at this problem, a sample selection method based on unstable cuts for the compression of large data sets is proposed in this paper. The extreme value is obtained at the interval endpoint for convex function, and therefore the endpoint degree of a sample is measured by making the unstable cuts of all attributes according to the basic property. The samples with higher endpoint degree are selected,and the calculation of the distance between the samples is avoided. The efficiency of the computation is improved without affecting the classification accuracy. The experimental results show a significant effect of the proposed algorithm on the compression for the large data set with high imbalance ratio and strong ability of anti-noise. 2016 Vol. 29 (9): 780-789 [Abstract] ( 596 ) [HTML 1KB] [ PDF 475KB] ( 869 ) 790 Energy Aware Task Scheduling Algorithm in Cloud Workflow System LI Xuejun, XU Jia, WANG Futian, ZHU Erzhou, WU Lei In the research on cloud workflow systems, the time efficiency optimization of the task execution is the emphasis. The energy consumption optimization of the task execution is often ignored. However, time-optimal task scheduling plans have different energy consumption. Therefore, how to solve energy-optimal task scheduling plans with time constraint are discussed in this paper. Firstly, the energy model of task execution is improved. Then, the fitness computation method of the task plan is designed to evaluate energy consumption. Finally, an adaptive inertia weight computation method is applied to adjust particle velocity accurately and a particle swarm optimization (PSO) algorithm is presented to solve the energy consumption optimization problem of task scheduling in cloud workflow systems. Experimental results show that the proposed algorithm has a stable convergence speed with low energy consumption. 2016 Vol. 29 (9): 790-796 [Abstract] ( 520 ) [HTML 1KB] [ PDF 538KB] ( 654 ) 797 Graph Cut Segmentation Method Based on Multiple Priori Shape XIN Yuelan , WANG Xili, ZHANG Xiaohua, HUANG Heming To segment multiple objects in the graph, a graph cut segmentation method for multiple priori shape constraints is proposed. The shape distance in a discrete level set framework is used to define the priori shape model ,and then this model is merged into the regional item of the graph cut framework. The priori energy function is expanded by adding multiple shape priors. The weight coefficient of shape prior item is adaptively adjusted to realize the adaptive control of shape items accounted for the proportion of the energy function. And thus, the problem of artificial selection of parameters is solved and the efficiency of segmentation is enhanced. To obtain the invariance of the method proposed in this research for shape affine transformation, the techniques combining the scale invariant feature transform and the random sample consensus are employed to align. The experimental results indicate that multiple targets in the image can be segmented by the proposed method. Moreover, the image noise pollution as well as occlusion is inhibited. 2016 Vol. 29 (9): 797-806 [Abstract] ( 710 ) [HTML 1KB] [ PDF 1823KB] ( 626 ) 807 Cascaded Hidden Space Fuzzy C-means Based on Local Preserving Projection LIU Huan, WANG Jun, YING Wenhao, WANG Shitong The traditional fuzzy clustering algorithms have poor learning ability for complex nonlinear data. Aiming at this problem, a condensed hidden space feature mapping is proposed by combining local preserving projection (LPP) and extreme learning machine (ELM) feature mapping. Thus, the original data is mapped into the condensed ELM hidden space. By connecting several condensed hidden space feature mapping together and combining fuzzy clustering methods, the cascaded ELM hidden space is constructed and a cascaded hidden space fuzzy clustering algorithm is proposed. Experimental results show that the proposed algorithm is insensitive to fuzzy index and efficient and robust for non-linear data and image data with intra-class variation. 2016 Vol. 29 (9): 807-815 [Abstract] ( 519 ) [HTML 1KB] [ PDF 474KB] ( 576 ) Researches and Applications 816 Image Automatic Segmentation Based on Fast Online Active Learning YAN Jing, PAN Chen, YIN Haibing An algorithm for image segmentation is proposed by building a pixels classification model. The model is trained online fast with a feed-forward neural network. Firstly, saliency map is computed by spectral residual (SR) approach. Then, multi-scale analysis is conducted via dispersion of minority high saliency points, and saliency map and gaze areas highly matching with human visual system are obtained. Next, positive and negative samples are selected randomly from saliency and non-saliency regions to compose the training set. A two-class random weighted feed-forward neural network model is trained. Finally, whole image pixels are classified by this model, and image segmentation is realized. Experiments show that the proposed algorithm enhances the segmentation performance for salient object grounded on the spectral residual based method, and the segmentation results are close to human visual perception. 2016 Vol. 29 (9): 816-824 [Abstract] ( 602 ) [HTML 1KB] [ PDF 1393KB] ( 791 ) 825 Learning Word Embeddings for Paraphrase Scoring in Knowledge Base Based Question Answering ZHAN Chendi, LING Zhenhua, DAI Lirong The conventional word embeddings are learned from the co-occurrence probabilities between the words within a same sentence. The learning algorithm is task-independent and unsupervised. A method for constructing word embeddings is proposed by utilizing the constraints of paraphrasing to improve the performance of paraphrase scoring with word embeddings and bag-of-words model in knowledge base (KB) based question answering (QA). In the proposed method, the pairs of paraphrase questions and non-paraphrase questions are collected respectively from a database of question paraphrases according to some designed rules. Then, the inequalities describing the similarities between the pairs of questions are adopted to represent the semantic constraint at the sentence level. These inequalities are integrated into the objective function for training word embeddings. Experimental results show that the proposed method improves the accuracies of paraphrase scoring and KB-based question answering compared with conventional word embedding methods. 2016 Vol. 29 (9): 825-831 [Abstract] ( 435 ) [HTML 1KB] [ PDF 417KB] ( 950 ) 832 Minimal Hepatic Encephalopathy Classification Based on Discriminative Subgraph Reconstruction TU Liyang, DU Junqiang, JIE Biao, ZHANG Daoqiang Minimal hepatic encephalopathy (MHE) is related to the abnormality of subnetworks, but searching related subnetworks is still a challenging task. To solve this problem, a method based on discriminative subgraph reconstruction is proposed to search subnetworks related to MHE and the subnetworks are used for MHE classification. Firstly, frequent subgraphs are mined from the functional connectivity networks of MHE and non-MHE (NMHE), respectively. Next, the discriminative subgraphs are selected from the frequent subgraphs for the original networks reconstruction and the combination of discriminative networks is conducted to reconstruct the original networks. Finally, the graph kernel is applied to compute the similarity between pairwise reconstructed networks and the kernel SVM is adopted for MHE classification. On the dataset of 77 patients with hepatic cirrhosis, the high accuracy of the proposed algorithm is achieved and the effectiveness of the proposed method is demonstrated. 2016 Vol. 29 (9): 832-839 [Abstract] ( 439 ) [HTML 1KB] [ PDF 639KB] ( 714 ) 840 Parallel Extreme Learning Machine Based on Improved Particle Swarm Optimization LI Wanhua, CHEN Yuzhong, GUO Kun, GUO Songrong, LIU Zhanghui To improve the stability of extreme learning machine(ELM), an extreme learning machine based on improved particle swarm optimization (IPSO-ELM) is proposed. By combining the improved particle swarm optimization with ELM, IPSO-ELM can find the optimal number of nodes in the hidden layer as well as the optimal input weights and hidden biases. Furthermore, a mutation operator is introduced into IPSO-ELM to enhance the diversity of swarm and improve the convergence speed of the random search process. Then, to handle the large-scale electrical load data, a parallel version of IPSO-ELM named PIPSO-ELM is implemented with the popular parallel computing framework Spark. Experimental results of real-life electrical load data show that PIPSO-ELM obtains better stability and scalability with higher efficiency in large-scale electrical load prediction. 2016 Vol. 29 (9): 840-849 [Abstract] ( 715 ) [HTML 1KB] [ PDF 576KB] ( 639 ) 850 Semi-supervised Classification Algorithm Based on l1-Norm and KNN Superposition Graph ZHANG Yunbin, ZHANG Chunmei, ZHOU Qianqian, DAI Mo A framework is proposed to construct a graph revealing the intrinsic structure of the data and improve the classification accuracy. In this framework, a l1-norm graph is constructed as the main graph and a k nearest neighbor (KNN) graph is constructed as auxiliary graph. Then, the l1-norm and KNN superposition (LNKNNS) graph is achieved as the weighted sum of the l1-norm graph and the KNN graph. The classification performance of LNKNNS-graph is compared with that of other graphs on USPS database, such as exp-weighted graph, KNNgraph, low rank graph and l1-norm graph, and 5% to 25% of the labeled samples are selected in experiments. Experimental results show that the classification accuracy of LNKNNS-graph algorithm is higher than that of other algorithms and the proposed framework is suitable for graph-based semi-supervised learning. 2016 Vol. 29 (9): 850-855 [Abstract] ( 615 ) [HTML 1KB] [ PDF 356KB] ( 766 ) 856 Convolutional Neural Network Algorithm Based on Double Optimization for Image Recognition LIU Wanjun, LIANG Xuejian, QU Haicheng To improve the recognition accuracy and the convergence speed of the convolutional neural network algorithm, a convolutional neural network algorithm based on double optimization is proposed. By modeling a convolutional neural network and optimizing the process of feature extraction and regression classification, an optimization convolutional neural network is built. Thus, the integrated optimization of the convolution and the full-connection process is realized. Compared with the local optimization network, the integrated optimization network obtains a higher convergence speed and better recognition accuracies. The experiments are conducted based on handwritten digit datasets and face datasets and the results show the improvement of the convergence speed and the recognition accuracy. And the effectiveness of the proposed algorithm is demonstrated. Moreover, this optimization strategy can be further extended into other deep learning algorithms related to convolution neural networks. 2016 Vol. 29 (9): 856-864 [Abstract] ( 827 ) [HTML 1KB] [ PDF 560KB] ( 1111 )

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