Design and implementation of improved CNN activation function

Abstract

Convolutional neural network has powerful feature learning capabilities and are widely used in the field of image classification. In this paper, an image classification method with improved CNN activation function is proposed. By analyzing the shallow convolutional neural network, a CIFAR-10 image classification model is constructed. In the process of data preprocessing, the digital standardization of the images is completed and the sample labels are one-hot encoded. The model network structure proposed in this paper adopts the ReLU nonlinear activation function and maximum pooling. The training results show the accuracy of the classification model is significantly improved. At the end of this paper, the accuracy rates of the four activation functions of Sigmoid, Tanh, ReLU, and T-ReLU are compared, and the advantages of the unsaturated nonlinear activation function are pointed out. The model is improved by using the T-ReLU activation function, with the accuracy rate increasing from 62% to 76.52%.