Enhanced image denoising via colour component decomposition and bit-plane specific CNN models

Image denoising is a fundamental task in computer vision aimed at restoring clean images from noisy inputs. Conventional methods often struggle with synthetic images due to the sensitivity of the human visual system to such noise. This paper proposes a novel Colour Component Decomposition Denoising Network (CCDDNet) for blind Gaussian image denoising, where noise levels are unknown during testing. The approach decomposes images into bit planes and denoises RGB components separately using dedicated convolutional neural networks with multi-layer residual learning. This progressive strategy effectively removes noise while preserving fine details and overall image quality. Experimental results on benchmark datasets, including BSD100, Kodak24, and Urban100, demonstrate superior performance, achieving up to 37.78 dB PSNR and 0.978 SSIM, outperforming state-of-the-art methods such as Restormer and KBNet. These results confirm the effectiveness and robustness of CCDDNet across varying noise levels and diverse image conditions.