CINEMO-NET: implicit neural and diffusion-based joint demosaicking and denoising

Abstract

Demosaicking and denoising of Colour Filter Array (CFA) images are challenging due to incomplete spectral sampling, sensor noise, zipper artefacts, and high-frequency detail loss. To address these limitations, this paper proposes a novel CFA-aware Implicit Neural Representation Diffusion-Integrated Memory Optimization Network (CINEMO-NET) for joint demosaicking and denoising. The framework incorporates a CFA-aware Implicit Neural Representation (CINR) module for continuous spatial modelling and missing-colour reconstruction, a Diffusion-Integrated Deep Memory Unrolled Optimization Network (D2MUO-Net) for reconstruction refinement and noise control, and a Score-Based Diffusion Proximal Denoising module for suppressing sensor noise and interpolation artefacts. A Fourier Neural Operator (FNO)-based Frequency Consistency Refinement module further preserves global spectral consistency and high-frequency details. Experiments on the McMaster Image Demosaicing Dataset with synthetic Gaussian noise demonstrate reconstruction-loss reductions of 60.41%, 68.85%, and 44.12% and SNR improvements of 25.17%, 17.76%, and 11.53% over DDNM, Multispectral Demosaicking, and CNN-Transformer approaches, respectively.