Non-local means algorithm with adaptive patch size and bandwidth

Non-local means algorithm is an effective denoising method that consists in some kind of averaging process carried on similar patches in a noisy image. Some internal parameters, such as patch size and bandwidth, strongly influence the performance of non-local means, but with the difficulty of tuning. Many solutions for choosing these two parameters, like cross-validation and Steins unbiased risk estimate criterion, are successful but computationally heavy. In this paper, we introduce a new feature metric that is capable of providing a quantitative measure of geometric structures of image in the presence of noise. The proposed region-based non-local means method first classifies a noisy image into several regions. Then, a local window and a local bandwidth value are selected pixel-wisely according to the property of each region and the local value of the new feature metric. Experiments on standard test images show that the proposed method outperforms the original non-local means version by around 1.34 dB and is comparable to or better than the performance of the current state-of-the-art non-local means based denoising algorithms, both visually and quantitatively.