Toward Blind Joint Demosaicing and Denoising of Raw Color Filter Array Data

Abstract Raw color-filter-array (CFA) data collected in the real world are often noisy and signal-dependent, which makes it difficult to recover the full-resolution noise-free color image. Denoising and demosaicing are two popular tools developed for noisy CFA data in modern color imaging pipeline. However, most existing works on joint demosaicing and denoising (JDD) are based on ad hoc assumptions about image degradation process; while in practice little is known about noise statistics (e.g., noise level) and processing pipeline (e.g., gamma correction). We advocate a blind formulation of joint demosaicing and denoising (bJDD) problem in this paper and present a novel divide-and-conquer approach toward blind reconstruction from noisy raw CFA data. Instead of making over-simplified assumptions about noise statistics, we propose to develop a more realistic Poisson-Gaussian noise model for simulating noisy raw CFA data in the real world. We also introduce a sub-network to adaptively estimate the noise level map from the noisy input, which will provide supplementary information to the deep model for non-blind JDD. Finally, we have adopted a generative adversarial network (GAN) based network for further perceptual optimization. Our extensive experimental results have shown convincingly improved performance over existing state-of-the-art methods in terms of both subjective and objective quality metrics.