The Research Of Image Classification Method Based On Convolutional Neural Network

The product display map of the store,the simulation map of the building,and the emoticons in the mobile chat,the images are everywhere.Image information is intuitive,vivid and easy to understand,and has become an indispensable way for people to communicate.In the times of big data,more and more images have increased the difficulty of obtaining the image which you need,and put forward higher requirements for image classification technology.With the rise of deep learning,convolutional neural networks is better than traditional image classification techniques in image classification,attracting some attention.The local receptive field of the convolutional neural network is similar to the way in which visual information is processed in the human brain.The convolution operation of the convolutional neural network fully considers the two-dimensional spatial structure of the image,and can extract the features of the image structure.The weight sharing and pooling operations of the convolutional neural network reduce the parameters in the model training,optimize the model,and speed up the convergence.Therefore,convolutional neural networks are suitable for dealing with image problems.The performance of the convolutional neural network depends largely on the network model structure and optimization algorithms.Based on this,this thesis studied the network model structure and optimization algorithms.The main contents are as follows:(1)In this thesis,thorough studied the characteristics,structure,calculation,optimization algorithm and training process of the convolutional neural network.Considered the differences of data sets,the advantages of convolutional neural network and optimization algorithms,this thesis constructed different convolutional neural networks for different data sets.(2)Learned the gradient descent algorithm and its two regularization methods:L2 regularization and weight decay.In the stochastic gradient descent algorithm(SGD)and Adam gradient descent algorithm,compared the difference of L2 regularization and weight decay in the process of accelerating convergence and preventing over-fitting.Theoretically,explained the reason why L2 regularization does not achieve optimization in the more complex gradient descent algorithm.Based on this,we improved the Nadam gradient descent algorithm,and proposed the Nadam gradient descent algorithm based on weight decay.(3)With TensorFlow as the back end,this thesis used the deep learning framework Keras to construct convolutional neural network,and combined with the proposed algorithm,the effectiveness of the algorithm is verified on the image classification problem.On the classic MNIST handwritten dataset and Cifar10 dataset,compared the experimental results of the proposed algorithm and other popular gradient descent algorithms.As the experimental results show,the proposed algorithm is better and the classification loss is lower than other popular gradient descent algorithms.