Color Image Demosaicking using a 3-stage Convolutional Neural Network Structure

ICIP 2018 Kai Cui, Zhi Jin, Eckehard Steinbach Chair of Media Technology, Technical University of Munich

Abstract

Color demosaicking (CDM) is a critical first step for the acquisition of high-quality RGB images with single chip cameras. Conventional CDM approaches are mostly based on interpolation schemes and hand-crafted image priors, which result in unpleasant visual artifacts in some cases. Motivated by the special characteristics of inter-channel correlations (higher correlations for R/G and G/B channels than that for R/B), in this paper, a 3-stage convolutional neural network (CNN) structure for CDM is proposed. In the first stage, the G channel is reconstructed independently. Then, by using the reconstructed G channel as guidance, the R and B channels are recovered in the second stage. Finally, high-quality RGB color images are reconstructed in the third stage. The objective and visual quality evaluation results show that the proposed structure achieves noticeable quality improvements in comparison to the state-of-the-art approaches.

Proposed Structure

Inspired by inter-channel correlation, in this paper, we propose a 3-stage CNN structure for color demosaicking.

CNNCDM-3Stage

Evaluation

You can download the original images and zoom in for better visual quality, the difference for area marked with the green box can be easily observed. (The 2-stage results are based on our reimplementation since the source code is not available yet, all other results are generated from the original source codes downloaded from the authors’ webpages.)