The Study Of Image Classification Algorothm Based On Deep Neural Network

With the rising of deep learning, the accuracy of image classification which is based on the deep neural network is significantly improved. Deep Learning makes the image classification technology achieve a new height. However the accuracy of image classification still needs to be further improved in application.In order to improve the classification accuracy of deep learning, based on the character of the random process of deep learning, this paper present a new joint way of network: based on jointing the difference features to compose a symmetric deep neural network, and apply it to image classification. The main contributions of this paper are as follows:Firstly, we propose a new method to combine different model which is based on the difference of feature extracted from the same image by different deep model, the symmetric model is constructed by combing the output layer of two different deep networks. Explains the reason why the different deep neural networks model can extract different feature from the same image, and verify the difference by experiments. Two sub-networks extracted the features by forward propagating separately in the symmetric network, in order to take advantage of the difference, we propose an ideal that joint the last feature layer in each sub-network together, and use the difference to optimize the loss function. Then we present the way of connecting the feature layer of subnetwork with class layer.Secondly, the fundamental of difference feature and the ensemble way can develop the original deep neural network into symmetric model. In this paper, we develop the deep belief network(DBN) into symmetric deep belief network(SDBN) develop the convolution neural network(CNN) into symmetric convolution neural network(SCNN). Design the constructing and training way of SDBN and SCNN, and we pose the method of measuring subnetworks' difference by analyzing the difference of subnetwork, the difference can be as penalty term to optimize the loss function, and then fine tune the model parameters by error back propagation to obtain much more accuracy in image classification task.On the Matlab platform, the experiments are conducted on MNIST and CIFAR-10 datasets, and the result demonstrates that the proposed method achieves improved accuracy compared with DBN and CNN architecture.