Learning to Joint Remosaic and Denoise in Quad Bayer CFA via Universal Multi-Scale Channel Attention Network

Abstract

The color filter array widely used in smart phones is mainly Quad Bayer and Bayer. Quad Bayer color filter array (QBC) is a filter shared by four pixels, which can improve the image quality by averaging four pixels in the 2 $\,\times \,$ × 2 neighborhood under low light conditions. From low-resolution Bayer to full-resolution Bayer has become a very challenging research, especially in the presence of noise. Considering denoise and remosaic, we propose a general two-stage framework JRD-QBC (Joint Remosaic and Denoise in Quad Bayer CFA), including denoise and remosaic. To begin with, for the denoise phase, in order to ensure the difference of each color channel recovery, we convert the input to hollow QBC, and then enter our backbone network, including source encoder module, feature refinement module and final prediction module. After that, get a clean QBC and then use the same network structure to remosaic to generate Bayer. Extensive experiments demonstrate the proposed two-stage method has a good effect in quantitative indicators and subjective vision.