NL-Means algorithm
Noise Gaussian noise priori based on the averaged non-local, proposed by the Baudes in 2005, the algorithm uses common natural image noise come redundant information. And conventional bilinear filtering, median filtering the image using local information to different filtering is that it utilizes the whole image denoising, an image block units to find similar regions in the image, then find these regions average, better able to remove Gaussian noise present in the image. NL-Means filtering process can be represented by the following formula:
BM3D算法(Block-matching and 3D filtering)
Idea of the algorithm is somewhat similar with the NL-Means is also looking for ways to block similar filtering in the image, but relative to the NL-Means much more complex understanding of the NL-Means algorithm helps to understand BM3D. BM3D algorithm There are two major steps, divided into basic estimate (Step1) and the final estimate (Step2):
In both stride, there are three small steps: grouping similar blocks (Grouping), collaborative filtering (Collaborative Filtering) and aggregation (Aggregation). Flowchart algorithm above will have relatively well represented this process, and only need a little explanation.
NL-Means and BM3D can be said is the best de-noising algorithm, which BM3D even claim that it can get by far the highest PSNR. We can see the final result, BM3D effect is indeed better than the NL-Means, less noise, better able to recover the details of the image. Wins in effect this point BM3D. Worthy of the State-of-the-art in this title. Of course, the test sample is relatively small, probably not enough to fully explain the problem.