Joint image demosaicking and denoising with mutual guidance of color channels

Abstract

Deep network based methods have shown the advantages in image demosaicking and denoising for raw images. However, how to fully utilize the correlations of color channels in raw data still needs further investigations. In this paper, we propose a novel end-to-end network for joint demosaicking and denoising (JDD) with mutual guidance of color channels (MGCC), which utilizes the strengths of all color channels comprehensively. Specifically, RGB channels are reconstructed by three branches respectively and each color channel can guide the reconstruction of others by a designed channel guiding block (CGB), which fuses features from different color channel reconstruction branches spatial-adaptively. Moreover, considering the amount of data in the additional channel, i.e., the extra green channel in RGGB Bayer pattern, features from its reconstruction branch are applied as guidance first. In this way, each color channel is capable of learning the advantages of other color channels adaptively in different regions to recover itself better. Through this mechanism, our method can achieve results with more details and less artifacts. Experimental results on both synthetic and realistic datasets demonstrate that our method outperforms state-of-the-art JDD methods both quantitatively and qualitatively. The code is publicly available on: https://github.com/yongzhangwhu/MGCC-JDD.