Joint Demosaicing and Denoising via Aliasing-Free Frequency Decomposition

Abstract

Joint demosaicing and denoising (JDD) remains challenging due to severe aliasing and color artifacts in regions with high-frequency details and strong color edges. Despite employing complex architectural designs to model inter-channel relationships or decouple the tasks, conventional methods still struggle to effectively evaluate spatial and inter-channel correlations. In this paper, we demonstrate that explicit frequency decomposition effectively decouples both spatial patterns and inter-channel relationships, providing a structured foundation for JDD. To prevent spectral leakage, we propose the Aliasing-Free Frequency Decomposition (AFFD) block along with a novel Kernel Frequency Regularization (KFR) loss that strictly enforces intended spectral band constraints. Furthermore, we introduce the Kernel Frequency Band Energy Ratio (K-FBER) to quantitatively verify the layer-level transfer function characteristics of kernels. Extensive experiments demonstrate that our proposed framework achieves state-of-the-art restoration quality with exceptional parameter efficiency, without relying on overly complex architectures or structures. Specifically, our large variant (FDJDD-L) utilizes less than half the parameters of the Transformer-based DemosaicFormer (14.63M vs. 34.25M) while gaining + 0.21 dB PSNR and + 0.0009 SSIM on the MIT Moiré dataset. Additionally, our compact FDJDD-M (7.32M) surpasses existing state-of-the-art methods by + 0.09 dB PSNR and + 0.0003 SSIM on the McMaster benchmark.