Optimization of CNN and Its Application in Handwritten Numeral Recognition

Abstract

Convolutional neural network (CNN) has been widely applied in the field of handwritten numeral recognition, but the discuss about the parameter of CNN network structure is pretty few. This paper made in-depth analysis of CNN network systematically and make in-depth research on network structure, the number of hidden feature map, the size of convolutional kernel, weight initialization, the number of bulk samples, iteration times and other parameters in recognition process and provide their value choice method. This paper also proposed a method based on CNN analysis and experiment and optimized the parameters of CNN aimed for MNIST handwritten numeral database. In addition, this paper proposed the process and method applied in handwritten numeral parameter optimization based on CNN image recognition, which has a good reference value for the further application of CNN network in the field of image recognition.