Research Of Convolutional Neural Network In Image Classification

Deep learning technology is an important branch of machine learning, it has gained great achievements in image and video analysis, speech processing and Natural Language Processing. In the field of image recognition, the widely used convolutional neural network is the focus of this paper.In order to improve the accuracy of image classification, this paper will study the deepconvolutional neural network. On the one hand,we will improve the activation function of deep convolution neural network, and enhance the network generalization ability and prevent overfitting.On the other hand, the combination of convolutional neural network model and support vector machine is used to construct a model for image classification. And in Ubuntu14.03 system,based on theano framework, with NVIDIA GTX750 Ti, the effectiveness of this method is verifiedby experiment.The main work of this paper as follows:(1) In this paper,we analyse the merits and defects of activation function Re Lu, Softplus. And based on them wef construct a piecewise function.By designing a convolutional neural networks,experiment on public data sets, analyzing various neuronal excitation function to the networkconvergence speed and the influence of image recognition accuracy.(2) Overfitting is a serious problem in the deep learning, Dropout is a way to prevent model from overfitting, and to improve the generalization ability of network. The method of max pooling with Dropout not only overcomes the shortcomings of the average pooling and max pooling, but also introduces the randomness, which is a robust method for network regularization. We improved the max pooling layer in the testing phase, and proposed a new method of probability weighted pooling, and the generalization ability was verified by experiments.(3) Under the condition of limited hardware, combined with modified activation function and the optimized network model, the convolutional neural network and support vectormachine(SVM) is connected to complete image recognition task.Convolution neural network is a kind of deep learning algorithm, for image feature extraction, and then in the last layer support vector machine(SVM)is used for classification.By experimental verification, image classification effect is achieved.