The Optimization And Application Of Convolution Neural Network For Image Recognition

Convolution Neural Network(CNN) is a kind of artificial Neural Network based on multilayer supervision learning network, and the problem of traditional pattern recognition method can be improved by CNN to extract features easily. It not only has the advantages of traditional neural network, good fault tolerance, adaptability and self-learning ability, but also has the ability of automatic feature extraction and the characteristic of weight sharing. So it has been widely used in the field of image recognition, object detection and recognition and target tracking, etc. In the process of convolutional neural network, the network structure optimization is an important factor to affect the accuracy and efficiency of the identification. It is important to study the structure optimization of CNN network.The theory, the characteristic and structure of CNN are analysed, and the basic process of training and testing in CNN is researched. The network structure, number of hidden layer characteristics figure, size of convolution kernel, initialization of weight, size of sample batch and number of iterations in the process of CNN recognition can be analysed, then parameter values can be given. In the process of recognition based on CNN, number of hidden layer characteristics figure has great influence on the accuracy. Recognition accuracy of the system can be affected by some small relational samples between a layer features figure and the next layer features figure. Internal relational between data is excavated by grey relational analysis, GRA is brought into the process of network training to automatically select effective features of hidden layers and optimize the network structure.Convolution kernel size and network structure layer can be determined by the size of the input image, and determines size of sample batch and number of iterations with using experimental method. At the same time, grey correlation analysis method is used to determine the number of hidden layer characteristic figure to determine the network structure and related parameters, handwritten numeral recognition and traffic sign recognition experiment can be finished by CNN with good structure.Finally, the number of hidden layer characteristic figure by grey correlation analysis method is according with the number of local optimal value by the experimental method and the number of characteristic figure can be adaptively determined by the proposed method and the optimization of CNN network structure can be completed, which can be shown by results.