Bayesian nonlocal means image denoising based on principal neighborhood dictionaries

Abstract

Nonlocal means (NLM) is an effective denoising filter. As an extension of NLM filter, Bayesian nonlocal (BNL) means filter provides a general framework adapted to different noise and is better parametrized than NLM filter. However, as processing in noisy image patches, the filter is not effective for large noise removal. Principal neighborhood dictionary (PND) based on principal component analysis (PCA) was proposed to achieve a high denoising accuracy. In this paper, we proposed a new BNL filter based on PND. Our filter applys the BNL framework to PCA subspace to improve the denoising results for noisy image with large standard deviation noise. Furthermore, according to different noise models, we present two filters for natural image denoising and synthetic aperture radar (SAR) image despeckling respectively. Experimental results tested on natural images and SAR images demonstrate that our filter reaches state-of-the-art performance both subjectively and objectively.