The Research On Deep Learning Algorithm And Its Application In Image Classification

Recently, there is growing concern about researches on Deep Learning(DL). Compared with the traditional shallow network models, the multi-layer network structure of deep learning model can express the complex function more effectively so that it can obtain more representative features. As a result, applying Deep Learning into image classification can effectively improve the accuracy of classification. However, Deep Learning still faces serious challenges in practical applications. Restricted Boltzmann Machine(RBM), an important basic unit model in deep learning, has high computation complexity of training process and low likelihood of training data which leads to a very long training time. In this thesis, the method of image classification based on deep network model is studied. The four main contributes are as follows:(1) The background and significance of Deep Learning are introduced. The history, basic structure units and the commonly used depth models and other related technologies of Deep Learning are described. The researches and development status of deep learning application are summarized. Existing problems of RBM are emphatically analyzed.(2) In order to solve the problem of slow convergence rate of traditional learning rate, a Kaiser Window function based RBM learning algorithm is proposed. With Kaiser Window, the algorithm can freely choose the characteristics of the proportion between the width of the main lobe and side lobe height as well as the variation of the error increment in the training process. In addition, the formula of adaptive learning rate is obtained, and the error control factor and inertia factor are set up. Thus the learning rate can be adjusted more effectively and adaptively. Simulations show that the proposed Kaiser Window function based RBM learning algorithm can effectively improve the convergence rate and learning ability of RBM.(3) In order to solve the problem of high computational complexity of the normalized parameters, an alternative iteration algorithm based RBM is proposed. This algorithm will alternately calculate distribution parameters and the normalization parameters of the model until they get convergent to obtain the distribution parameters of RBM. At the same time, it will set the corresponding threshold to ensure the convergence rate. Simulations show that, the improved RBM based on alternative iteration algorithm has higher likelihood compared with the RBM models obtained by using parallel tempering or persistent contrastive divergence.(4) In order to solve the problem that shallow layer model can not effectively express the image semantic features of multilayer, the Deep Learning is involved in the image classification and an image calssification method based on multilayer RBM network is proposed. And the process of adding noise and the pooling process are also involved in the multilayer RBM network. As a result, the over fitting on training data can be avoided and generalization ability and robustness of the classification method can be improved. Compared With the deep belief network, convolutional neural network and sparse encoding image classification method, simulations show that the proposed image classification method based on multi layers RBM network can improve the image classification accuracy as well as the model's generalization and robustness.