模式识别与人工智能
Home      About Journal      Editorial Board      Instructions      Ethics Statement      Contact Us                   中文
Pattern Recognition and Artificial Intelligence
22 Judgement and Disposal of Academic Misconduct Article
22 Copyright Transfer Agreement
22 Proof of Confidentiality
22 Requirements for Electronic Version
More....
22 Chinese Association of Automation
22 National ResearchCenter for Intelligent Computing System
22 Institute of Intelligent Machines,Chinese Academy of Sciences
More....
 
 
2019 Vol.32 Issue.4, Published 2019-04-25
Papers and Reports    Ressarches and Applications   
   
Papers and Reports
289 Collaborative Filtering Recommendation Algorithm Based on Time-Related Correlation Degree and Covering Degree
ZHANG Zhipeng, ZHANG Yao, REN Yonggong
The traditional item-based collaborative filtering(IBCF) assigns equal weights to all items while computing similarity and prediction. And thus it cannot provide recommendations with high predictive accuracy and classification accuracy. Therefore, a time and covering weighting collaborative filtering(TCWCF) algorithm is proposed. A time-related correlation degree is applied to similarity computation to improve the predictive accuracy, and a covering degree is integrated into rating prediction to increase classification accuracy. Experimental results on MovieLens dataset suggest that TCWCF outperforms traditional IBCF and other algorithms and it provides recommendations with satisfactory predictive accuracy and classification accuracy for users.
2019 Vol. 32 (4): 289-297 [Abstract] ( 600 ) [HTML 1KB] [ PDF 771KB] ( 309 )
298 Online Video Stabilization Algorithm on Lie Group Manifold
YANG Jiali, LAI Linjing, ZHANG Lei, HUANG Hua
The traditional video stabilization algorithms cannot achieve good performance with a low latency. Aiming at this problem, an algorithm is proposed using the Lie group manifold based Kalman filter to stabilize videos in real time. The video frame motion is separated into rotation component and translation component. The rotation component can be represented by the rotation matrix obtained via the gyroscope data, while the translation component is provided by the translation matrix computed by the matching between video frames. Both the sequence of the rotation matrices and the translation matrices can form the motion paths on the Lie group manifold. Therefore, the Lie group manifold based Kalman filter is used to smooth the rotation and translation component, respectively. Finally, the video frame sequence can be stabilized by the motion compensation. The experimental results show that the proposed algorithm achieves high online real-time video stabilization performance on the mobile device.
2019 Vol. 32 (4): 298-305 [Abstract] ( 485 ) [HTML 1KB] [ PDF 1127KB] ( 289 )
306 Method of Online Learning Resource Recommendation Based on Multi-objective Optimization Strategy
LI Haojun, YANG Lin, ZHANG Pengwei
Single-objective transformation method is commonly used in online learning resource recommendation. In the recommendation process, the consideration of learner preference is inadequate. Therefore, the accuracy of learning resource recommendation is affected. An online learning resource recommendation model, multi-objective resource recommendation model(MOSRAM), is proposed based on multi-objective optimization strategy. In this model, learner preference for the type of learning resources and the fitness of the difficulty level are regarded as the optimization objectives in the planning time. A multi-objective particle swarm optimization algorithm, neighborhood multi-objective particle swarm optimization(NEMOPSO), with the ability to benefit from neighbor mean and explore new regions is designed. An online learning resource recommendation method, neighborhood multi-objective particle swarm optimization-resource recommendation approach(NEMOPSO-RA), based on MOSRAM model is proposed. The comparison of recommendation methods with classical multi-objective optimization algorithms under different problem scales show that the accuracy and performance of online learning resource recommendation can be effectively improved by NEMOPSO-RA method.
2019 Vol. 32 (4): 306-316 [Abstract] ( 419 ) [HTML 1KB] [ PDF 946KB] ( 330 )
317 Semi-supervised Network Representation Learning Model Based on Graph Convolutional Networks and Auto Encoder
WANG Jie, ZHANG Xihuang
Combining graph convolutional networks(GCN) and auto encoder(AE), a scalable semi-supervised network representation learning model, Semi-GCNAE, is proposed to preserve the network structure information and node feature information. GCN is utilized to capture the structure and feature information of all nodes in K-order neighborhood of the node. The captured information is utilized as the input of AE. The K-order neighborhood information captured by GCN is extracted and the dimension is reduced nonlinearly by AE. The cluster structure information of nodes is preserved by combining Laplacian feature mapping. The ensemble learning method is introduced to train GCN and AE jointly. Therefore, the learned low-dimensional vector representation of nodes can retain both network structure information and node feature information. Extensive evaluation on five real datasets shows that the low-dimensional vector representation of nodes acquired by the proposed model preserves the structure and characteristics of the network effectively. And it generates better performance in node classification, visualization and network reconstruction tasks.
2019 Vol. 32 (4): 317-325 [Abstract] ( 582 ) [HTML 1KB] [ PDF 4742KB] ( 547 )
326 Visual Clustering Method of Quasi-Circular Mapping Based on Dimension Extension and Rearrangement
HUANG Shan, LI Ming, CHEN Hao, LI Junhua, ZHANG Congxuan
The non-linear structure of high-dimensional data cannot be captured by the existing radial layout visualization method. Therefore, visual clustering method of quasi-circular mapping based on dimension extension and rearrangement is proposed. The dimension of high-dimensional data is expanded by affinity propagation clustering algorithm and multi-objective clustering visualization evaluation index. Then, the dimension correlation rearrangement of the extended high-dimensional data is carried out. Finally, the high-dimensional data is reduced to two-dimensional visualization space by quasi-circular mapping mechanism to realize effective visual clustering. Experiments show that the proposed dimension extension and rearrangement strategy can effectively improve the visual clustering effect of quasi-circular mapping visualization. The dimension extension strategy can also significantly improve the clustering effect of other radial layout visualization methods with better generalization performance. Moreover, the proposed method has obvious advantages in visual clustering accuracy, topology
2019 Vol. 32 (4): 326-335 [Abstract] ( 344 ) [HTML 1KB] [ PDF 3455KB] ( 225 )
336 Fine-Grained Visual Classification Based on Sparse Bilinear Convolutional Neural Network
MA Li, WANG Yongxiong
The overfitting problem of bilinear convolutional neural network(B-CNN) for fine-grained visual recognition is caused by the large number of parameters and its complex structure. In this paper, a sparse B-CNN is proposed to handle the problem. Firstly, a scaling factor is introduced into each feature channel of B-CNN, and regularization of sparsity is applied to the scaling factors during the training. Then, the feature channels in B-CNN with low contribution to the final classification are identified by small scaling factors. Finally, these channels are pruned in a certain proportion to prevent overfitting and increase the significance of key features. The learning of sparse B-CNN is weakly supervised and end-to-end. The verification experiments on FGVC-aircraft, Stanford dogs and Stanford cars fine-grained image datasets show that the accuracy of sparse B-CNN is higher than that of the original B-CNN. Moreover, compared with other advanced algorithms for fine-grained visual recognition, the performance of sparse B-CNN is same or even better.
2019 Vol. 32 (4): 336-344 [Abstract] ( 444 ) [HTML 1KB] [ PDF 766KB] ( 607 )
Ressarches and Applications
345 Data Driven Optimal Stabilization Control and Simulation Based on Reinforcement Learning
LU Chaolun, LI Yongqiang, FENG Yuanjing
Q-learning algorithm is used to solve the optimal stabilization control problem while only the data, rather than the model of the plant, is available. Due to the continuity of state space and control space, Q-learning can only be implemented in an approximate manner. Therefore, the proposed approximate Q-learning algorithm can obtain only one suboptimal controller. Although the obtained controller is suboptimal, the simulation shows that the closed-loop domain of attraction of the proposed algorithm is broader and the cost function is also smaller than the linear quadratic regulator and deep deterministic policy gradient method for the strongly nonlinear plant.
2019 Vol. 32 (4): 345-352 [Abstract] ( 512 ) [HTML 1KB] [ PDF 1204KB] ( 575 )
353 Chip Image Super-Resolution Reconstruction Based on Deep Learning
FAN Mingming, CHI Yuan, ZHANG Mingjin, LI Yunsong

Since the convolutional neural networks can introduce the prior knowledge of the chip image in the training stage, a chip image super-resolution algorithm is proposed. A convolutional neural network is utilized to improve the initial reconstruction image of the iterative method, the complementary information between image sequences is employed through an iterative process and a chip sample set is built. Experimental results show that the proposed method produces clearer chip images with close packing and yields higher average values of the objective evaluation indicators. Furthermore, the proposed algorithm performs well on nature images.
2019 Vol. 32 (4): 353-360 [Abstract] ( 543 ) [HTML 1KB] [ PDF 1516KB] ( 482 )
361 Sketch Recognition Combining Deep Learning and Semantic Tree
ZHAO Peng, FENG Chencheng, HAN Li, JI Xia
In the existing sketch recognition based on deep learning, a whole sketch is employed as an input of network, and therefore the recognition process is uninterpretable. The semantic tree is introduced into sketch recognition based on deep learning, and a sketch recognition method, sketch-semantic net, is proposed in this paper. Firstly, data-driven segmentation method is utilized to divide a whole sketch into component sketches with the semantic information. Secondly, the component sketches are recognized by transfer deep learning. Finally, the component sketches are associated with the sketch categories according to the semantic concepts of the semantic tree, and thus the gap between low level semantics and high level semantics is reduced. The experimental results on the popular Sketch_ dataset demonstrate the effectiveness of the proposed method.
2019 Vol. 32 (4): 361-368 [Abstract] ( 451 ) [HTML 1KB] [ PDF 858KB] ( 283 )
369 Character-Based Disconnected Recurrent Neural Network for Name Nationality Identification
ZHANG Yusha, ZHANG Liming, JIANG Shengyi
Personal name is viewed as a strong indicator of inferring the nationality of the user. Generally, personal names reveal the differentiation and correlation of naming conventions among different nationalities. In the current research, personal name features are extracted by cutting off name strings into a set of independent n-gram units, while subtle relationships between characters are not explored. Therefore, a character-based disconnected recurrent neural network is proposed to capture subtle features among personal names in this paper. Concretely, a set of fragments is derived from name strings by order using a slice window. Then, long short-term memory units are utilized to learn information of each fragment, and they are aggregated via mean-pooling operation to obtain the whole name representation for nationalities prediction of users. Disconnected fragments enable model to focus on subtle features among different personal names. Experiments on Olympic dataset and Aminer dataset show that the proposed model outperforms the existing models and the performance is satisfactory.
2019 Vol. 32 (4): 369-375 [Abstract] ( 437 ) [HTML 1KB] [ PDF 697KB] ( 283 )
376 Human Action Recognition Based on Multi-view Semi-supervised Learning
TANG Chao, WANG Wenjian, WANG Xiaofeng, ZHANG Chen, ZOU Le
Since human action is complicated in nature, single action feature view lacks the ability of comprehensively profiling human action. A method for human action recognition based on multi-view semi-supervised learning is proposed in this paper. Firstly, a method based on three different modal views is proposed to represent human action, namely Fourier descriptor feature view based on RGB modal data, spatial and temporal interest point feature view based on depth modal data and joints projection distribution feature view based on joints modal data. Secondly, multi-view semi-supervised learning framework is utilized for modeling. The complementary information provided by different views is utilized to ensure better classification accuracy with a small amount of labeled data and a large amount of unlabeled data. The classifier-level fusion technology is employed to combine the predictive ability of three views, and the problem of confidence evaluation of unlabeled samples is effectively solved. The
2019 Vol. 32 (4): 376-384 [Abstract] ( 646 ) [HTML 1KB] [ PDF 806KB] ( 377 )
模式识别与人工智能
 

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