Multi-residuals Network and Region Constraints Based Face-image Denoising

Abstract

In recent years, the denoising models based on convolutional neural network (CNN) have made great progress. However, CNN based image denoising models tend to generate artifacts and blurry edges. To deal with this problem, this paper proposes a multi-residuals network with cascade strategy to keep image textures, and integrates face region constraints to loss function of model optimization. The weighted loss function characterizes the location and gray probabilities of different face regions, which brings benefits to recover face-image sharpness and naturalness. Experimental results on the Helen and IMM face datasets show that the proposed model can suppress artifacts in smooth regions and recover sharper edges.