ISSN 1000-1239 CN 11-1777/TP


    Default Latest Most Read
    Please wait a minute...
    For Selected: Toggle Thumbnails
    A Novel Regularization Method Based on Convolution Neural Network
    Lü Guohao, Luo Siwei, Huang Yaping, Jiang Xinlan
    Journal of Computer Research and Development    2014, 51 (9): 1891-1900.   DOI: 10.7544/issn1000-1239.2014.20140266
    Abstract3447)   HTML8)    PDF (3024KB)(2713)       Save
    Regularization method is widely used in solving the inverse problem. An accurate regularization model plays the most important part in solving the inverse problem. The energy constraints should be different for the different types of images and different parts of the same image, but the traditional L1 and L2 models used in the field of image restoration are both based on a single prior assumption. In this paper, according to the defects of the single priori assumption in traditional regularization model, a novel regularization method based on convolution neural network is proposed and applied to image restoration, therefore, the image restoration can be regarded as a classification issue. In this method, the image is partitioned into several blocks, and the convolution neural network is used to extract and classify the features of sub-block images; then the different forms of the priori regularization constraints are adopted considering the different features of the sub-block images, therefore the regularization method is no longer limited to a single priori assumption. Experiments show that the image restoration results by the regularization method based on convolution neural network are superior to those by the traditional regularization model with a single priori assumption. Therefore the regularization method based on convolution neural network can better restore image, maintain the edge texture characteristic of the image nicely, and has lower computational cost.
    Related Articles | Metrics
    A Sparse Stochastic Algorithm with O(1/T) Convergence Rate
    Jiang Jiyuan,Xia Liang,Zhang Xian, Tao Qing
    Journal of Computer Research and Development    2014, 51 (9): 1901-1910.   DOI: 10.7544/issn1000-1239.2014.20140161
    Abstract1559)   HTML1)    PDF (2898KB)(922)       Save
    Stochastic gradient descent (SGD) is a simple but efficient method for large-scale optimization problems. Recent researches have shown that its convergence rate can be effectively improved by using the so-called α-suffix averaging technique in solving the strongly convex problems. However, SGD is purely a black-box method, so it is difficult to obtain the expected effect on the learning structure while solving the regularized optimization problems. On the other hand, composite objective mirror descent (COMID) in the stochastic setting is a scalable algorithm which can effectively keep the sparsity imposed by L1 regularization;But it can only obtain an O(logT/T) convergence rate for the strongly convex optimization problems. In this paper, we focus on the generally convex optimization problem in the form of “L1 + Hinge”. To make full use of the α-suffix averaging technique, we first change it into a strongly convex optimization problem by adding an L2 strongly convex term. Then, we present an L1MD-α algorithm which combines the COMID algorithm and the α-suffix averaging technique. We prove that the L1MD-α algorithm can achieve an O(1/T) convergence rate. As a result, our L1MD-α algorithm not only obtains faster convergence rate but also better sparsity than COMID. Through extensive experiments on some typical large-scale datasets we finally verify the correctness of the theoretical analysis and effectiveness of the proposed algorithm.
    Related Articles | Metrics
    Remote Sensing Image Classification Based on DBN Model
    Lü Qi, Dou Yong, Niu Xin, Xu Jiaqing, Xia Fei
    Journal of Computer Research and Development    2014, 51 (9): 1911-1918.   DOI: 10.7544/issn1000-1239.2014.20140199
    Abstract2816)   HTML14)    PDF (3333KB)(2300)       Save
    Remote sensing image classification is one of the key technologies in geographic information system (GIS), and it plays an important role in modern urban planning and management. In the field of machine learning, deep learning is springing up in recent years. By mimicking the hierarchical structure of human brain, deep learning can extract features from lower level to higher level gradually, and distill the spatio-temporal regularizes of input data, thus improve the classification performance. Deep belief network (DBN) is a widely investigated and deployed deep learning model. It combines the advantages of unsupervised and supervised learning, and can archive good classification performance for high-dimensional data. In this paper, a remote sensing image classification method based on DBN model is proposed. This is one of the first attempts to apply deep learning approach to urban detailed classification. Six-day high-resolution RADARSAT-2 polarimetric synthetic aperture radar (SAR) data were used for evaluation. Experimental results show that the proposed method can outperform SVM (support vector machine) and traditional neural network (NN).
    Related Articles | Metrics
    Image Classification Using Hierarchical Feature Learning Method Combined with Image Saliency
    Zhu Jun,Zhao Jieyu,Dong Zhenyu
    Journal of Computer Research and Development    2014, 51 (9): 1919-1928.   DOI: 10.7544/issn1000-1239.2014.20140138
    Abstract1525)   HTML6)    PDF (3754KB)(1175)       Save
    Efficient feature representations for images are essential in many computer vision tasks. In this paper, a hierarchical feature representation combined with image saliency is proposed based on the theory of visual saliency and deep learning, which builds a feature hierarchy layer-by-layer. Each feature learning layer is composed of three parts: sparse coding, saliency max pooling and contrast normalization. To speed up the sparse coding process, we propose batch orthogonal matching pursuit which differs from the traditional method. The salient information is introduced into the image sparse representation, which compresses the feature representation and strengthens the semantic information of the feature representation. Simultaneously, contrast normalization effectively reduces the impact of local variations in illumination and foreground-background contrast, and enhances the robustness of the feature representation. Instead of using hand-crafted descriptors, our model learns an effective image representation directly from images in an unsupervised data-driven manner. The final image classification is implemented with a linear SVM classifier using the learned image representation. We compare our method with many state-of-the-art algorithms including convolutional deep belief networks, SIFT based single layer or multi-layer sparse coding methods, and some kernel based feature learning approaches. The experimental results on two commonly used benchmark datasets Caltech 101 and Caltech 256 show that our method consistently and significantly improves the performance.
    Related Articles | Metrics
    Sparseness Representation Model for Defect Detection and Its Application
    Li Qingyong,Liang Zhengping,Huang Yaping, Shi Zhongzhi
    Journal of Computer Research and Development    2014, 51 (9): 1929-1935.   DOI: 10.7544/issn1000-1239.2014.20140153
    Abstract1210)   HTML2)    PDF (1471KB)(1009)       Save
    Defect detection is an important applicaion of computer vision in industry, but it is a challenge to effectively inspect defects in a vision system because of illumination inequality and the variation of reflection property of objects. Sparseness is one of the most improtant characteristics of defect images, and therefore the approach named defect decomposition based on sparseness (DDBS) is proposed. In the framework of DDBS, a defect image is treated as linearly combination of three components: background, defects and noise. Background is coded by an over-complete sparse dictionary, which can not sparsely represent defect component. Meanwhile defect is sparsely coded by its dictionary that is exclusive to background. Then defect component is decomposed using DDBS based on the principle of blind sources sepration and block-coordinate relaxation. Furthermore, DDBS is applied in rail surface defect detection to improve the robustness of inspection systems. Experiments are carried out for different dictionary combinations based on actual rail images, and the results demonstrate that DDBS can effectively detect the defects that are missed by the state-of-the-art methods. DDBS is a flexible framwork for applications of defect detection and has the potential benefit to improve robustness of traditional methods.
    Related Articles | Metrics
    A Study of Speech Recognition Based on RNN-RBM Language Model
    Li Yaxiong, Zhang Jianqiang, Pan Deng, Hu Dan4
    Journal of Computer Research and Development    2014, 51 (9): 1936-1944.   DOI: 10.7544/issn1000-1239.2014.20140211
    Abstract2725)   HTML11)    PDF (1524KB)(1391)       Save
    In the recent years, deep learning is emerging as a new way of multilayer neural networks and back propagation training. Its application in the field of language model, such as restricted Boltzmann machine language model, gets good results. This language model based on neural network can assess the probability of the next word appears according to the word sequence which is mapped to a continuous space. This language model can solve the problem of sparse data. Besides, some scholars are constructing language model making use of recurrent neural network mode in order to make full use of the preceding text to predict the next words. From these models we can sort out the restriction of long-distance dependency in language. This paper attempts to catch the long-distance information based on RNN-RBM. On the other hand, the dynamic adjunction of language model ia analyzed and illustrated according to the language features. The experimental result manifests there are considerable improvement to the efficiency of expanding vocabulary continuing speech recognition using RNN_RBM language model.
    Related Articles | Metrics
    Cited: Baidu(8)
    Audio Classical Composer Identification by Deep Neural Network
    Hu Zhen, Fu Kun, Zhang Changshui
    Journal of Computer Research and Development    2014, 51 (9): 1945-1954.   DOI: 10.7544/issn1000-1239.2014.20140189
    Abstract1867)   HTML6)    PDF (2371KB)(2055)       Save
    Music is a kind of signal that has hierarchical structure. In music information retrieval (MIR) area, higher level features, such as emotion and genre, are typically extracted based on lower level features such as pitch and spectrum energy. Deep neural networks have good capacity of hierarchical feature learning, which indicates that deep learning is potentially to obtain good performance on music dataset. Audio classical composer identification (ACC) is an important problem in MIR which aims at identifying the composer for audio classical music clips. In this work, a hybrid model based on deep belief network (DBN) and stacked denoising autoencoder (SDA) is built to identify the composer from audio signal. The model get an accuracy of 76.26% in the testing data set which is better than some thoroughbred models and shallow models. After dimensionally reduced by linear discriminant analysis (LDA) it is also clear that the samples from different classes become farther away from each other when being transformed by more layers in our model. By comparing models in different sizes we give some empirical instruction for ACC problem. Similar to image, music features are hierarchical too and different parts of our brain handle signals differently. So we propose a hybrid model and our results encourage us to believe that our proposed model makes sense in some applications. During the experiments, we also find some practical guides for choosing network parameters. For example, number of neurons in the first hidden layer should be approximately 3 times to the dimension of input data.
    Related Articles | Metrics