Region characteristics-based fusion of spatial and transform domain image denoising methods

Abstract

Nonlocal means (NLM)- and wavelet-based image denoising methods have drawn much attention in imageprocessing due to their effectiveness and simplicity. The performance of these algorithms varies according to regioncharacteristics in an image. For example, NLM performs well for smooth regions due to deployment of redundancyavailable in images, whereas wavelet-based approaches may preserve key image features by controlling the degree ofthreshold for shrinking the noisy coefficients. This paper presents a simple novel approach that estimates an original imageby simply taking the weighted average of the denoised images pixel values obtained by NLM and wavelet thresholdingschemes based on natural characteristics of regions in an image. Extensive simulations on standard images demonstratethat the proposed approach outperforms the benchmark wavelet-based schemes, NLM and its variants, in terms of peaksignal-to-noise ratio (PSNR(dB)), mean structural similarity metric (MSSIM), and visual quality