Image steganalysis with multi-scale residual network

In recent years, many deep neural network models are used in steganalysis. However, the deep neural network models on steganalysis usually use the single scale channel for detection. When the number of convolution kernels reaches a certain limit, the improvement of detection accuracy is very weak by increasing the number of convolution kernels. In this paper, we try to establish a wider range of image region correlation extraction, and propose a multi-scale deep neural network model. The model is based on the deep residual network and adopts end-to-end design. Different local receptive fields in the same layer were selected to generate the characteristic channels. By the channel recognition, variety of image steganographic features were achieved from different scale channels. Experiments show that the multi-scale residual network can further improve the accuracy of steganography detection more than the networks of the single scale channel.