Enhanced Image Denoising with Diffusion Probability and Dictionary Learning Adaptation

Abstract

Image denoising is essential for numerous image processing applications, where image noise can profoundly impact processing efficiency and output quality. Addressing the challenge of inflexible reference images in unconditional diffusion probability models and enhancing image denoising performance is of paramount importance. In this research, we propose a novel image denoising model based on component decoupling and introduce sensitivity decoupling operators to prevent entanglement and redundancy among different decoupling models. Additionally, we leverage a model-driven network to fuse image components, resisting noise and model degradation, thereby aiding network convergence. Subsequently, we construct an image adaptive denoising model incorporating diffusion probability and dictionary learning. Experimental results demonstrate the superiority of the proposed approach over other algorithms in grayscale image processing on the Set12 dataset, achieving a peak signal-to-noise ratio (PSNR) of 35.75 dB and an average structural similarity (SSIM) value of 92.68%. Similarly, on the BSD68 dataset, our algorithm outperforms others with a PSNR of 34.35 dB and an average SSIM of 93.89%. Furthermore, for colour image processing, our method yields higher PSNR and average SSIM compared to other approaches. The findings indicate a significant improvement in denoising effectiveness compared to prior methods, highlighting the practical value of the proposed image denoising algorithm.