Image denoising via nonlocal means with noise-robust similarity

We propose an effective image denoising method by revising the conventional “nonlocal means (NL-means)” method. Conventional NL-means replaces a noisy pixel by the weighted average of other reference pixels depending on the similarity between local neighborhoods of target pixel and reference pixels. Noise, however, reduces the similarity even within the same pattern blocks. This is due to the weighted average of dissimilar blocks of pixels, which makes the result image blurred or irregular. Hence, our proposal is to calculate the similarity with the combination of basis patterns correlated statistically with little noise through principal component analysis making use of local structures that have low correlation with incurred noise. The first step is to exclude the local structures of the given image whose statistical correlation with the original images is slight. The second step is to exclude basis patterns whose correlation with the target blocks is slight. Making use of the selected basis patterns, we can calculate the similarity robust to any noise. Consequently, high-quality images can be obtained. Comparing with other methods, our experiments show our method denoises effectively without causing any blur or irregularity.