The Analysis of CNN Structure for Image Denoising
Jae Hyeon Park, Jeong Hyeon Kim, Sung In Cho · 2018
This paper proposes an optimal structure of a convolutional neural network (CNN) for image denoising by analyzing the conventional CNN denoisers. There are three main factors that can determine the denoising performance of the CNN denoiser: the number of feature dimensions of each convolution layer, the number of convolution layers, and the usage of dilated convolution. We analyze the denoising performance variations of the conventional CNN denoiser depending on the above three factors and propose the optimal structure of the CNN denoiser. Experimental results showed that the above three factors have a high correlation with the denoising performance. Based on the experimental results, we could provide the optimal structure of the CNN denoiser.