Nonlocal-Means Denoising Algorithm Research And Its GPU Implementation

The nonlocal-means denoising (NLM) algorithm explicitly exploits self-similarities in images. It is one of the most powerful denoising algorithms. But sometimes it is also too powerful to protect the edges, so it is not suited for images with much texture. The computation of NLM is very large, which prevent it from practical application. This paper is mainly about the improved NLM algorithm suited for texture images and its GPU parallel implementation.This paper has two contributions on NLM method.(1) We combined the NLM algorithm with the steering kernel in kernel regression (KR) method, and introduced SK-NLM(steering kernel nonlocal-means) algorithm. Since the original NLM method uses Gaussian kernel so that the edges will be over-smoothed. Experimental results showed that replacing the Gaussian kernel with steering kernel will preserve the edges more efficiently.(2) Introducing the approximate k-nearest neighbors matching to the NLM method, we called it AKNN-NLM method. It finds K neighbors by randomly searching in a more large areas and use K-nearest neighbors to denoise. This paper also makes the parallel optimization on AKNN-NLM and implemented it on NVIDIA Geforce GT430GPU by CUDA. Experimental results showed that it gained a maximally19acceleration rate in denoising.