Image Classification Based On Deep Learning

As the results of the computer technology's development,great changes have taken place in people's way of life.What used to be done by people can now be done by machine.Which have been influenced most is image classification.Traditional classification is mainly described based on the visual features of images and requires manual adjustment algorithm.However,with the development of deep learning,features can be extracted by machines,which can learning and summarize the basis of classification,just like the human brain.In this paper,the framework of TensorFlow is used to build the network,and the network is designed to distinguish the pictures of different dog species.In the model training,the network design is mainly according to the pricinple of convolutional neural network,and finally get a successful results.Then,the Inception structure is added to the network to compare and analyze the changes in the accuracy rate of different network structures.The specific work and achievements of this paper are as follows:1.The development status and future trend of deep learning are analyzed.Introducing the origin of deep learning and its relationship with neural networks Describing the principles of multi-layer neural networks,cyclic neural networks and convolutional neural networks.Moreover,formula calculation is made for gradient problems and using Adam optimization methods in neural network.Finally,convolutional neural network is selected as the basis for the network construction in this paper.2.A dog species recognition algorithm based on convolutional neural network is proposed.According to the traditional convolutional neural network,the activation layer and the Local Response normalized layer are added to enhance the model expression.Dropout technology is used to reduce randomly discarded neurons and alleviate the possible of overfitting problems in training.At the same time,Adam optimization algorithm was added,and combined with the cross entropy and learning rate decay to optimize the parameters in the process of back propagation to make the model converge faster,and the accuracy reached 92.25% at last.3.Proposing the classification and recognition algorithm based on Inception struction.The flipping operations are used to amplify the sample data set and whiten the image to remove the influence of surroundings factors,reduce the probability of over-fitting.On the basis of the convolutional neural network structure proposed above,the Inception structure is added to increase the width and depth of the network,and the final accuracy rate is about 93.25%.