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.8, Published 2015-08-30 Papers and Reports    Researches and Applications        Papers and Reports 673 Restricted Boltzmann Machine Based Spectrum Modeling and Unit Selection Speech Synthesis Method SONG Yang, LING Zhen-Hua, DAI Li-Rong A restricted Boltzmann machine based spectrum modeling and unit selection speech synthesis method is proposed. At the model training stage, the restricted Boltzmann machine is used to model spectral features with rich details, such as spectral envelopes and short-time spectral amplitudes, instead of using the single Gaussian model with diagonal variance and mel-cepstrum feature for spectral model in the traditional approach. Thus, the description capability of the acoustical model for spectral feature is improved. At the speech synthesis stage, the restricted Boltzmann machine model is adopted to calculate the log likelihoods of spectral feature of candidate sample, and a method of piecewise linear mapping is proposed to construct target cost function for unit selection. The experimental results indicate that the proposed method can effectively improve the naturalness of synthetic speech. 2015 Vol. 28 (8): 673-679 [Abstract] ( 569 ) [HTML 1KB] [ PDF 467KB] ( 647 ) 680 Multi-label Emotion Classification Based on Decision-Theoretic Rough Set ZHANG Zhi-Fei, MIAO Duo-Qian, ZHANG Hong-Yun To solve the problem of multi-lable uncertainty in emotion classification, a multi-label classification method based on decision-theoretic rough set, named DTRS-MLC, is proposed. The positive, negative, and boundary regions with the multi-label mapping function are defined by the dual-weighted multi-label K-nearest neighbor (DW-ML-KNN) algorithm, and the label co-occurrence and label exclusiveness relationship with the label dependency degree is described. From the perspective of theoretical and experimental analysis of the relationship between DTRS-MLC and DW-ML-KNN, DW-ML-KNN can be viewed as a special case of DTRS-MLC. The experimental results on music and text emotion classification tasks show that DTRS-MLC achieves better performance as a whole. 2015 Vol. 28 (8): 680-685 [Abstract] ( 522 ) [HTML 1KB] [ PDF 392KB] ( 427 ) 686 Logistic Regression Method for Class Imbalance Problem GUO Hua-Ping, DONG Ya-Dong, WU Chang-An, FAN Ming As one of the most important classification models in pattern recognition and machine learning, logistic regression(LR) is an interpretable model and has good generalization ability. In this paper, LR model is applied to class imbalance problem, and a method, named LR for class imbalance (LRCI), is proposed to tackle data imbalance problem. To take a full consideration of data imbalance, two objective functions g-mean based metric (FBM) and f-measure based metric(GBM) are constructed respectively to supervise LRCI learning model parameters. And then, the model is effectively quaranteed high accuracy and recall rate. The experimental results on UCI datasets show that LRCI significantly boosts the performance on recall, g-mean and f-measure in the premise of high accuracy of LRCI. Besides, LRCI presents significant advantage comparing to other state-of-the-art class imbalance learning model. 2015 Vol. 28 (8): 686-693 [Abstract] ( 572 ) [HTML 1KB] [ PDF 397KB] ( 754 ) 694 Rapid Registration Algorithm of Large-Scale Images Based on Normalized Gradient Phase Correlation CHEN Huai-Yu, YANG Yang To solve the real-time registration problem of large-scale images with rotations, scalings, translations simultaneously, an image registration algorithm based on normalized gradient phase correlation is proposed in this paper. The complicated multilayer computation, interpolation and iteration is avoided in this algorithm. Plural gradient images are disposed by normalized gradient phase correlation.Giving consideration to robustness and rapidity of parameters estimation at the same time, this algorithm can efficiently expand the estimation range of transformation parameters. By means of parameter-adjustable window function, it can suppress the influence of the edge effect of the different kinds of images. Experimental results illustrate the rapidity and effectiveness of the proposed algorithm. 2015 Vol. 28 (8): 694-701 [Abstract] ( 542 ) [HTML 1KB] [ PDF 878KB] ( 657 ) 702 3D Wavelet Octcube Split Video Coding Algorithm WANG Xiang-Hai, FU Ming-Zhe, SONG Chuan-Ming Video coding based on 3D wavelet transform draws more and more attention. However, it is necessary to improve its coding efficiency. Based on the analysis of the 3D wavelet transform coefficients distribution characteristics of video and unavailability of zerotree coding in temporal high-frequency frames, the zerocube model of temporal high-frequency frames and the 3D wavelet video coding algorithm based on octcube split are proposed in this paper. In this algorithm, low-frequency subbands and high-frequency subbands obtained by 3D wavelet transform are processed respectively, the zerotree structure of set partitioning in Hierarchical tree (SPIHT) is adopted for the former, the zerocube structure of octcube split is utilized for the latter. Since the 3D transform wavelet coefficients are organized together by zerocube structure, the synchronous information of lonely zero coding is reduced. Meanwhile, a coefficient is coded only when it is significant. Thus, to scan insignificant coefficients repeatedly is avoided. Experimental results show that for various standard test video sequences, the proposed algorithm has good reconstruction quality. 2015 Vol. 28 (8): 702-709 [Abstract] ( 388 ) [HTML 1KB] [ PDF 912KB] ( 482 ) 710 Higher Order Tensor Diffusion Magnetic Resonance Sparse ImagingBased on Compressed Sensing FENG Yuan-Jing, WU Ye, ZHANG Gui-Jun, LIANG Rong-Hua High Order Tensor (HOT) diffusion magnetic resonance imaging is an important method to reveal the microstructural information of living brain white matter. However, its time-consuming data acquisition and low fiber reconstructing resolution limit its clinical application.In this paper, a fiber orientation estimation method with weighted sparse is presented based on HOT model. Firstly, the HOT spherical deconvolution model for fiber orientation estimation is established. Then, a sparse representation method of fiber orientation is put forward. Finally, a l1-norm optimization model is built for the sparse constraint deconvolution. A computing method of applying the training results of sparse dictionary with lower order into the high-order problem is proposed to achieve the solution of the optimization problem. The experimental results on simulated and vivo data show that the fiber orientation estimation method improves the angular resolution of HOT tensor imaging method and reduces the angle recognition error. 2015 Vol. 28 (8): 710-719 [Abstract] ( 610 ) [HTML 1KB] [ PDF 2742KB] ( 704 ) Researches and Applications 720 Detection Method of Trace Multi-component Gases Based on SVM and PCA YU Dao-Yang, QI Gong-Mei, QU Ding-Jun, LI Min-Qiang, LIU Jin-Huai Gas sensors and optical sensors are difficult to detect trace multi-component gases. In this paper, a detection method of fast chromatography combined with gas sensor array is introduced to obtain the characteristic signal of trace multi-component gases. Support vector machine (SVM) is introduced to classify the samples according to the features. Then, to obtain a better gas identification model, particle swarm optimization algorithm is utilized to optimize the parameters of SVM. Based on actual sample detection and recognition, and compared to detection method by similar testing instrument, the proposed method has better selectivity of mixed gases. Using SVM、 PCA and PSO method is more suitable for processing and identification of small sample data. Developed multi-component gas detection prototype has better recognition rate, repeatability and stability. 2015 Vol. 28 (8): 720-727 [Abstract] ( 515 ) [HTML 1KB] [ PDF 755KB] ( 611 ) 728 Gene Expression Data Clustering Based on Projection Least Square Regression Subspace Segmentation JIAN Cai-Ren, CHEN Xiao-Yun Subspace segmentation method is an important method for machine learning. The existing researches on subspace segmentation method are generally on the original sample space. Advanced by existing dimensional reduction methods, a gene expression data clustering method based on projection subspace segmentation is proposed by joining projection method and least square regression based subspace segmentation. Projection matrix and remodeling matrix is got by using alternate optimization, and dimension reduction and cluster is realized simultaneously. The experimental results on six gene expression datasets illustrate the validity of the proposed method. 2015 Vol. 28 (8): 728-734 [Abstract] ( 422 ) [HTML 1KB] [ PDF 893KB] ( 530 ) 735 Pseudo-Relevance Feedback Technique Based on Matrix Factorization ZHOU Dong, LIU Jian-Xun, ZHANG San-Rong The performance of pseudo-relevance feedback technique is heavily dependent on two parameter values. Under the lack of relevance valuation results, these parameters can only rely on experience to set. In this paper, a pseudo-relevance feedback technique based on matrix factorization is proposed. This technique fuses multiple pseudo-relevance feedback results using the ideas of collaborative filtering together. And the optimal parameters are automatically selected for query expansion. Experimental results show that compared with the existing pseudo-relevance feedback techniques, the proposed method has a better retrieval performance, regardless of any underlying information retrieval model. 2015 Vol. 28 (8): 735-740 [Abstract] ( 538 ) [HTML 1KB] [ PDF 432KB] ( 582 ) 741 Multi-granulation Dominance Relation and Information Fusion Based on Set-Valued Information System ZHUANG Ying, LIU Wen-Qi, FAN Min, LI Jin-Hai In recent years, many kinds of multi-granulation rough set model have developed. Aimed at the dominance relation of set-valued information system which widely exists in practise, the multi-granulation dominance relation rough set model is proposed in this paper. Through the example of data processing, and by the comparison with the existing multi-granulation rough set model, the lower approximation of the multi-granulation dominance relation rough set is more accurate. Finally, the proposed model as a kind of information fusion technology is applied in the problem of target threat assessment. Experimental results show that the proposed model adapts more complex multi-source information environment. 2015 Vol. 28 (8): 741-749 [Abstract] ( 462 ) [HTML 1KB] [ PDF 410KB] ( 431 ) 750 Three-Dimensional Active Appearance Model Based on Tensor Mode andIts Application in Lung Field Segmentation from CT Images WANG Qing-Zhu, SUN Na , WANG Bin In the existing 3D active appearance model (AAM), a 3D appearance matrix must be transformed into a 1D vector. Thus, the segmentation accuracy is affected by the transformation and the segmentation efficiency is affected by the generated oversize vector. For the above problems, a 3D tensor based AAM is proposed aiming at direct operating on 3D appearances of lungs by higher order singular value decomposition, rather than transforming it from 3D to 1D. Firstly, the tensor-based model is built and the parameters are deduced. Then, a block-based Kronecker is designed to determine the optimal scheme for low rank representation of appearance tensors. This optimization avoids repeated computation. Finally, the whole segmentation system is constructed and used in the segmentation of lung images. Experimental results of clinical CT images are compared to other 3D landmark-based methods. The proposed model performs a better result in both precision and efficiency. 2015 Vol. 28 (8): 750-759 [Abstract] ( 357 ) [HTML 1KB] [ PDF 1461KB] ( 712 ) 760 Image Object Localization Based on Multiple Image Segmentation Scoring SHEN Xiang-Jun, MU Lei, ZHA Zheng-Jun, GOU Jian-Ping, ZHAN Yong-Zhao Image objects localization detects object regions accurately and therefore it can improve accuracies of image objects recognition and classification. In this paper, an image object localization method based on multiple segmentation regions scoring is proposed. Through the image of the multi-level segmentation, the semantic constraints among different image regions through multi-level segmentation results is confirmed. By these constraints, frequent itemset mining and scoring strategy are applied on different levels of object region pattern. According to pattern scores of regions, important regions in each segmentation levels are merged successively to localize the whole image object region. Experimental results on MSRC and GRAZ datasets show that the proposed method can localize image foreground object accurately, and its validity is verified on Caltech256 dataset. 2015 Vol. 28 (8): 760-768 [Abstract] ( 450 ) [HTML 1KB] [ PDF 666KB] ( 495 )

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