Curvelet Thresholding with Multiscale NLM Filtering for Color Image Denoising

This paper presents a multichannel Curvelet based image denoising scheme using hard thresholding and multiscale Non-Local Means (NLM) filtering. For the suppression of ringing artifacts due to hard thresholding and better localization of local structures like: edges, textures and small details the reconstructed image is further processed using Guided Image Filter (GIF). Each channel of the image is decomposed in three different scales including approximation and the finest scale. The use of NLM filter in the approximation and finest scale removes both the coarser grain (low frequency) and fine grain (oscillatory) noise independently in different channels. Hard thresholding in the remaining coarser scale separates the signal from noise effectively. Experimental results on TID2008 image database demonstrate the competitiveness of proposed denoising technique in terms of PSNR and SSIM at lower noise strength and excels in performance at higher noise level compared to several state-of-the-art algorithms including BM3D.