Optimized Color Filter Array for Denoising Diffusion Null-Space Model-Based Demosaicing

Recently, deep learning-based demosaicing methods have shown promising results. However, there has been little research on designing CFAs that are well-suited for specific deep learning-based demosaicing methods. This is because it is challenging to establish a relationship between deep learning-based demosaicing methods and the CFAs they employ. This contrasts with traditional CFA design methods, which targeted fixed, non-deep learning demosaicing methods that did not depend on data learning, making signal processing theory applicable to the design. In this paper, we propose an optimized color filter array(CFA) tailored for the Denoising Diffusion Null-space Model (DDNM) based demosaicing. We begin by demonstrating the application of the DDNM to the demosaicing problem and establish the conditions under which the DDNM can accurately recover the true colors from CFA images containing colored pixels. Based on this analysis, we propose a CFA pattern that significantly improves the likelihood of accurate color reconstruction using the DDNM-based demosaicing method. Then, we outline the training process for obtaining the optimal filter coefficient values for the proposed CFA pattern. Experimental findings demonstrate that the proposed CFA yields favorable results when paired with the DDNM-based demosaicing technique which surpass those achieved by other CFA patterns.