GMSD-based Perceptually Motivated Non-local Means Filter for Image Denoising

Due to increasing proliferation of multimedia signals, specifically, image, video and their applications in our daily life, it is indispensable to have methods that can efficiently predict and correct visual quality of images with high measures of accuracy. Therefore, in this work a state-of-the-art (STOA) image quality assessment (IQA) metric, gradient magnitude similarity deviation (GMSD) has been incorporated in a STOA least-square-based non-local means (NLM) filtering framework for image denoising. The denoising process works by estimating and weighting neighbouring patches similar to the patch being denoised in terms of Euclidean distance (ED) and GMSD coefficient. The overall process is broken down into two steps; initially, local noise estimates for the underlying noisy patch are approximated and removed, then the refined patch is fed to the weighting process as the final step. Further, the proposed methodology also helps in mitigating the patch jittering blur effect (PJBE) and over smoothing of denoised images as observed with conventional NLM algorithm. Experimental evaluations based on visual-quality assessment and least-square-based metrics have shown that the proposed algorithm yields better denoised image estimates than the conventional NLM algorithm. Moreover, experiments conducted on a subjective database, i.e. CSIQ, have shown higher performance in terms of peak signal-to-noise ratio (PSNR), structural similarity (SSIM) and GMSD coefficients. The resultant denoised images were in high correlation with the subjective judgements compared to the ones obtained with conventional NLM algorithm.