|M.Sc Student||Hila Berkovich|
|Subject||Model-Based Adaptive Non-Local Means Image Denoising|
|Department||Department of Electrical Engineering||Supervisors||Professor Emeritus Malah David|
|Dr. Bar-Zohar Meir|
|Full Thesis text|
Image denoising is used to find the best estimate of the original image given its noisy version. Among the vast image denoising methods that were suggested, patch-based approaches have drawn much attention in the image processing community. The Non-Local Means (NLM) denoising algorithm, first introduced by Buades et al. in 2005, takes advantage of image redundancy by comparing pixel neighborhoods within an extended search region in the image. Each pixel value is estimated as a weighted average of all other pixels in this search region. These pixels are each assigned a weight that is proportional to the similarity between the local neighborhood of the Pixel of Interest (POI) and their local neighborhood, such that pixels with a similar neighborhood are assigned higher weights. The NLM denoising approach originally refers to Additive White Gaussian Noise (AWGN).
The participation of dissimilar pixels, which may be included in the extended search region, in the weighted averaging process, degrades the denoising performance. To eliminate their effect, researchers suggest creating an adaptive search region that excludes those pixels. These suggested methods are parameter dependent and involve heuristics.
In this thesis, we present a novel model-based method that extracts a set of similar pixels for a given POI from its initial search region, using the statistical distribution of the NLM dissimilarity measure. Our approach does not require any parameter setting and provides better results than other compared adaptive search region approaches. Our scheme was also compared to the standard NLM and was found to provide better performance both quantitatively and visually.
We have also explored the effect of correlation between the dissimilarity elements of a given search region. Three sources of correlation were explored according to the degree of overlap between patches within a given search region. We found that the correlation-based model denoising results are comparable to those without it. Consequently, the much simpler model that discards the correlation is preferred.
The model-based scheme was also integrated in the Block-Matching 3D (BM3D) state-of-the art denoising scheme, such that the computational complexity of the original BM3D is reduced while denoising results remain comparable.
Besides the AWGN, we have explored our approach on Poisson noisy images as well. Poisson noise, which is signal dependent, is the noise type that characterizes images taken by a digital camera. We have verified that the tendency that characterizes the AWGN, both for adaptive model-based NLM and model-based BM3D, is preserved.