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....

2019 Vol.32 Issue.9, Published 2019-09-25 Papers and Reports    Researches and Applications        Papers and Reports 773 Two Stage Domain Adaptation Learning TIAN Lei, TANG Yongqiang, ZHANG Wensheng Aiming at minimization of the joint distribution difference between source domain and target domain in domain adaptation, a two-stage domain adaptation learning method is proposed. In the first stage, the discriminative information of sample labels and the data structure are considered, and a shared projection transformation is learned to minimize the difference of marginal distribution in the shared-projected space. In the second stage, an adaptive classifier with structural risk is learned by the labeled source data and unlabeled target data. The classifier minimizes the difference of conditional distribution of source domain and target domain as well as maintains the manifold consistency underlying the marginal distributions. Experiments on three benchmark datasets show that the method achieves better results on average classification accuracy and the Kappa coefficient. 2019 Vol. 32 (9): 773-784 [Abstract] ( 859 ) [HTML 1KB] [ PDF 706KB] ( 645 ) 785 Feature Extraction of Deep Topic Model for Multi-label Text Classification CHEN Wenshi, LIU Xinhui, LU Mingyu Traditional single-label feature extraction methods cannot effectively solve the problem of multi-label text classification. Aiming at this problem, a dual model of latent dirichlet allocation(LDA) and long short-term memory(LSTM), deep topic feature extraction model(DTFEM), is proposed in this paper. LDA and LSTM are employed as two channels, respectively. LDA is used to model global features of the text, and LSTM is used to model local features of the text. DTFEM can express the global and local features of the text simultaneously and combine supervised learning and unsupervised learning effectively to realize the feature extraction of different levels of text. Experimental results show that DTFEM is superior to other traditional text feature extraction models and obviously improves the indicators of multi-label text classification tasks. 2019 Vol. 32 (9): 785-792 [Abstract] ( 697 ) [HTML 1KB] [ PDF 644KB] ( 744 ) 793 Link Prediction Method Based on PU Learning LI Qi, WANG Zhiqiang, LIANG Jiye In classification-based link prediction methods, it is difficult to choose reliable negative examples to construct the link prediction classifier due to the large-scale and uncertainty of the unknown node pairs. Therefore, a link prediction method based on positive and unlabeled(PU) learning is proposed. Firstly, topological information of node pairs is extracted to construct example sets. Secondly, distribution of candidate negative examples is determined by community structure, and several candidate negative example sets are obtained through multiple under-sampling based on the distribution. Then, the classifiers constructed from multiple negative example sets and positive example sets are integrated to select reliable negative examples. Finally, the link prediction classifier is constructed based on positive examples and reliable negative examples. Experiments on four datasets show that the proposed link prediction method produces better prediction results than other related methods. 2019 Vol. 32 (9): 793-799 [Abstract] ( 484 ) [HTML 1KB] [ PDF 590KB] ( 352 ) 800 Multi-objective Evolutionary Algorithm Based on IGD+S LI Ming, DUAN Ruru, CHEN Hao, XIE Huihua How to evaluate solutions effectively is a key to solving many-objective optimization problem. An inverted generation distance(IGD+S) indicator is proposed based on IGD indicator, incorporating weak dominance of IGD+ indicator and employing the concept of non-contributing individuals. Convergence and diversity of solution set are evaluated comprehensively. IGD+S indicator is embedded in the framework of evolutionary algorithms, and a multi-objective evolutionary algorithm based on IGD+S indicator is presented. In the process of environmental selection, excellent solutions are selected according to enhanced IGD+S indicator. Experimental results demonstrate that the proposed algorithm is competitive in DTLZ problems and WFG problems. 2019 Vol. 32 (9): 800-810 [Abstract] ( 535 ) [HTML 1KB] [ PDF 855KB] ( 378 ) 811 Large-Scale Hierarchical Classification Online Streaming Feature Selection Based on Neighborhood Rough Set BAI Shengxing, LIN Yaojin, WANG Chenxi, CHEN Shengyu Label space of data possesses a hierarchical structure, and feature space is unknown and evolutionary in many classification learning tasks. An online streaming feature selection framework for large-scale hierarchical classification task is proposed. Firstly, a neighborhood rough model is defined for hierarchical structure data. Important features are dynamically selected based on feature correlation. Finally, the redundant dynamic features are identified based on feature redundancy. Experiments are conducted to verify the effectiveness of the proposed algorithm. 2019 Vol. 32 (9): 811-820 [Abstract] ( 511 ) [HTML 1KB] [ PDF 734KB] ( 361 ) Researches and Applications 821 Cross-View Gait Recognition Combined with Non-local and Part-level Features FENG Shiling, WANG Xiuhui In most existing gait recognition methods based on deep learning, global features are acquired by stacking convolutional layers, and local features beneficial to fine-grained classification are ignored. Aiming at this problem, a cross-view gait recognition method is proposed by combining non-local and part-level features. A pair of gait energy images(GEIs) is used as input to extract the non-local information of a single sample and the relative non-local information of the sample pairs. Then, human body regions are divided horizontally into static blocks, micro-dynamic blocks and strong dynamic blocks to extract better local features according to the geometric characteristics of GEI. Furthermore, the segmented regions are connected to three binary classifiers for training respectively. Finally, experiments on OU-ISIR-LP and CASIA-B gait datasets show that the proposed method produces a higher correct recognition rate. 2019 Vol. 32 (9): 821-827 [Abstract] ( 358 ) [HTML 1KB] [ PDF 565KB] ( 357 ) 828 Multi-aspect Sentiment Attention Modeling for Sentiment Classification of Educational Big Data ZHAI Guanlin, YANG Yan, WANG Heng, DU Shengdong Aiming at inefficiency and heavy workloads of college curriculum evaluation methods, a multi-aspect sentiment attention modeling(multi-ASAM) is proposed. Multi-ASAM concatenates a sentence and various aspects of the sentence by neural networks and adds emotional resources attention. To achieve better classification results, influence of relationships between aspects on emotinal polarity and contribution of emotional resources to emotional polarity is taken into auount in multi-ASAM. Experimental results show that Multi-ASAM is improved compared with other methods in the application of education and other fields. 2019 Vol. 32 (9): 828-834 [Abstract] ( 520 ) [HTML 1KB] [ PDF 685KB] ( 515 ) 835 Calligraphy Characters Dynamic Reproduction Algorithm Based on Principal Curve YANG Chenxu, ZHANG Hongyun, MIAO Duoqian Reproducing dynamic writing process of calligraphy works in an effective way is an urgent problem to be solved. Therefore, a calligraphy characters dynamic reproduction algorithm based on principal curve is proposed to guarantee the continuity and accuracy of strokes and solve the problem of many invalid branches caused by the thinning algorithm during skeleton extraction process. Based on the structure of calligraphy characters, the principal curve algorithm is improved and optimized. Firstly, skeleton extraction, skeleton tracking, stroke order acquisition, etc., are conducted. Then, the skeleton information and binary image are combined to realize stroke width restoration. Aiming at the tumor problem in the stroke width, the corner elimination method is proposed. Finally, the Canvas drawing technology of HTML5 is introduced into the dynamic writing process of calligraphy characters. Experiments on simulated datasets show that the proposed algorithm produces good results. 2019 Vol. 32 (9): 835-843 [Abstract] ( 448 ) [HTML 1KB] [ PDF 1561KB] ( 566 ) 844 Layout Adjustable Simulated Generation Method for Chinese Landscape Paintings Based on CGAN GU Yang, CHEN Zhaojiong, CHEN Can, YE Dongyi Creating a complete landscape painting via computer simulation is difficult without studying from global layout viewpoint. To address this issue, a layout-guided Chinese landscape painting simulation method for a complete painting generation is proposed. The characteristics of landscape paintings are taken into account in the design of feasible structures of layout label maps. Composition forms and elements of landscape paintings can be depicted using those structures. On the basis of condition generative adversarial network (CGAN) approach, a multi-scale feature fusion CGAN (MSFF-CGAN) is designed based on layouts and touches of landscape paintings. The proposed network is trained to accomplish heterogeneous transfer from a layout label map to a simulated landscape painting. To deal with rare availability of layout label maps for network training, a color pixel clustering algorithm with semantic correlation is used. In order to enhance the artistic reality of the generated landscape painting, a super resolution network named MemNet is incorporated to refine the texture details. Experimental results show that the proposed method is superior to existing methods in both integrity and artistic reality. Moreover, the proposed method can be used to handle simple graffiti sketches and modify simulated landscape paintings by editing label maps. 2019 Vol. 32 (9): 844-854 [Abstract] ( 582 ) [HTML 1KB] [ PDF 4909KB] ( 548 ) 855 Robust Face Sketch Synthesis Based on Deep Probabilistic Graphical Models ZHANG Yuqian, GAO Fangyuan, WANG Nannan Since pixel level features in the data-driven face sketch synthesis algorithms lack robustness to illumination variation and complex background, the quality of synthesized face sketches is poor. In this paper, a robust face sketch synthesis algorithm based on deep probabilistic graphical models is proposed. A preprocessing method is adopted to adjust illumination brightness and face pose of an input photo to make them consistent with the training photos. Instead of pixel feature, deep feature representation is utilized for neighbor selecting. A deep probabilistic graphical model is employed to jointly model the weight of sketch reconstruction and the weight of deep feature, and therefore the best reconstruction representation of the synthetic image is obtained. A fast nearest neighbor search method is proposed to speed up sketch synthesis. Experimental results verify robustness and rapidity of the proposed algorithm. 2019 Vol. 32 (9): 855-866 [Abstract] ( 476 ) [HTML 1KB] [ PDF 2724KB] ( 384 ) 867 2019 Vol. 32 (9): 867-867 [Abstract] ( 243 ) [HTML 1KB] [ PDF 189KB] ( 283 ) 868 2019 Vol. 32 (9): 868-868 [Abstract] ( 298 ) [HTML 1KB] [ PDF 188KB] ( 478 )

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