Image classification method based on improved deep residual networks

摘要

In order to solve the problem of image classification, a novel image classification method based on Residual Networks(ResNet) is proposed. Firstly, the 7*7 convolutional layer of the first layer of the network is replaced by a consequent three layer 3*3 convolutional layer, which reduces the number of model parameters without changing the receptive field. Secondly, the pooling layer of the network and the fully connected layer are replaced by the global average pooling layer, makes the model easier to train. Thirdly, the RelU function replaced by the better activation function Leaky ReLU. Finally, the model is verified by using crop disease images, and the experimental results show that the improved algorithm proposed in this study can effectively solve the problem of overfitting, and the classification of crop disease images reaches more than 98.3%, which is 1% higher than that of the original network.