Structure-Texture-Noise Decomposition for Noisy Images with Two-Stage Network

Structure-texture decomposition refers to the de-composition of an input image into three components: a structure component consisting of edges and smooth surface, a texture component consisting of local patterns. Most existing methods do not take the existence of the noise into account, but in practice, the noise is often included in the input image. In this study, we propose a deep learning based structure-texture-noise decomposition method that enables texture-noise separation using the context of structure components by sequentially connected two-stage network. Experimental results show that the proposed method can decompose noisy images into structure, texture, and noise components. Furthermore, we show that the proposed method can be applied to the tone mapping application with noisy input.