Research On Image Recognition Algorithm Based On Convolutional Neural Network

With the continuously improvement of computer science and the wider application field of image recognition,the research of image recognition based on deep learning method has developed rapidly.However,there are still some problems with convolutional neural network(CNN)such as long training time,slow convergence speed in the modeling stage,low recognition accuracy and poor robustness in the test stage.All those difficulties have become the bottleneck of CNN effective application in image recognition field.Therefore,the thesis has carried out theoretical and experimental research on the image recognition algorithm based on CNN,and the main work as follows:(1)In response of the problem of long training time of CNN,a rapid training method based on transfer learning is proposed.The method is used to the pre-training model which trained to solve other image recognition problem for transfer training,make the network quickly converge,so as to achieve the goal of shortening training time.The Experimental results of the MNIST and CIFAR-10 data sets show that the transfer learning model can indeed greatly improve modeling speed.(2)In response of the problem that the image feature extraction ability of maximum pooling and average pooling are not strong,which leads to low accuracy,the pooling method with Dropout is proposed.Combined with Dropout's randomness,the pooling operation process inhibits the expression of certain units according to probability and enhances the characteristic commonality training.Experimental results on MNIST and CIFAR-10 data sets show that the accuracy of the method is better than maximum pooling and average pooling method.(3)In response of the problem that activation function is prone to gradient disappearance and inability to train in practice,a new activation function is proposed.The function combines the slow gradient change of the Sigmoid function,the positive and negative output characteristics of the Tanh function,and the linear characteristics of the Re LU(Recitified Linear Unit)function.The experimental results show that the new activation function can avoid the problem of gradient disappearance and improve the training effect of the network.