The Research Of Non-Local Means Denoising Algorithm

As a method of image preprocessing, image denoising plays an important role in many applications related to image processing. So far, the scholars have successively put forward a lot of good denoising algorithm, non-local means (NLM) denoising algorithm is one of them. It is worth mentioning that, compared to the traditional filtering method, which only in the local neighborhood weighted average, the non-local means denoising algorithm can better preserve image edge structure information.The NLM algorithm makes use of the self-similarity information which widely exist in images for a weighted average. However, the block matching of NLM only consider the translation of the patch, it cannot handle rotation or mirroring, this shortage leads to the denoising performance still not good enough. So, in order to solve this problem, we propose two improved denoising schemes.The major work and our innovation can be summarized as follows:1) Directed against the problem in block matching, we propose a more accurate block matching method which considers the contribution of rotation. Given a pixel, we obtain the similar blocks to the center block(that is, the neighbor of the pixel) as follows:first, the center block is reordered according to their intensity values, and then the candidate blocks are also reordered in similar way, finally the distance between the center block and each candidate block is calculated according to the reordered intensity values of them. Only candidate blocks with small distance are selected. Furthermore, we obtain the more similar blocks structurally from the above candidate blocks:rotate the center block and each candidate block several times, and use their average values to compute the similarity between them. Finally, in order to eliminate impact of noise, we deal with the inputted image by a pre-filtering operation before the similarity calculation.2) We propose a rotation-invariant and noise-resistant similarity measure based on improved LBP operator, and use it as an alternative to the block matching algorithm in non-local means denoising algorithm. Before calculating the distance between two patches, we first obtain the improved LBP value of each patch and record the circular step of the binary string, then, rotating the patch based on the circular step to the same dominant orientation. In addition, in order to speed up the algorithm, an automatic selection strategy of similar patches is proposed.Experimental results demonstrate that the proposed method achieved higher peak signal-to-noise ratio (PSNR) and the mean structural similarity than some state-of-art methods. And our results were also more visual pleasing.