Multi-stage dense CNN demosaicking with downsampling and re-indexing structure

Abstract

Multi-stage CNN is composed of several networks and loss functions. Compared to a normal single-stage CNN, a multi-stage CNN can control the processing in each network. In image demosaicking, it is utilized to follow the flow of conventional non-CNN demosaicking algorithms. This paper proposes a 4-stage CNN demosaicking method. The first stage obtains a pre-estimated full color image from an input Bayer image instead of non-CNN pre-processing. The second stage processes the color planes independently while the third stage processes the combination for two colors. The final stage obtains the full-color outputs from all outputs of the former stages. To achieve better performance, the proposed method introduces dense structures for both inside and outside of every network. Another contribution is dealing with the discontinuity between a pixel and its neighboring pixels in Bayer images by a downsampling and re-indexing structure for feature maps. These contributions lead to higher accuracy and prevent false colors. Some results through demosaicking experiments show that the proposed method can achieve better quality images with less false colors and better PSNR than state-of-the-art CNN demosaicking methods.