Entropy-based Image Noise Variance Estimation And De-noising

In the process of image acquisition, transmission and storage there are always all kinds of noises being introduced, so the analysis of noise and processing is a typical problem of image processing. There are many de-noising algorithms, whose performance is well, such as the non-local average and BM3D, etc. On the one hand, these algorithms, assume that the level of noise image known, while actually the noise level of natural images is unknown, and the noise level of de-noising plays a key role. On the other hand, it is time consuming to find similar blocks, so it is hard for us to apply it to actual projects.This thesis studies this typical problem for image de-noising, and has obtained a noise variance estimation algorithm and an adaptive image de-noising method which are based on the noise variance estimation. The main work of this thesis covers the following contributions:First, Estimation algorithm of image noise:Based on the smoothing block, the noise estimated method is simple and fast. However, in the presence of noise, it is difficult to accurately locate the image smoothing block according to the variance. To solve this problem, this paper, according to the variance and entropy of image blocks, constructed a new evaluation mode to select smooth block, and introduce the image quality evaluation for further iterative filtering to select a good results; Based on the minimum heap of binary search to speed up the iterative, it can ensures that the estimation precision at the same time improve the processing speed of the algorithm.Second, Improved WNNM de-noising algorithms:WNNM de-noising, which is based on the local image block, the iterative parameter estimation is not accurate and time-consuming about block-matching. Based on the proposed noise variance, we first estimated globally optimal de-noising algorithm as WNNM initial parameters; second iteration parameter of iterative process is based on each local image block and its own calculation of the input and output, so the error is large. The weighted nuclear norm minimization de-noising algorithms at each iteration of each partial image blocks are required to enter a noise variance parameters, so we used the estimated adjacent regions optimal parameters to replace the idea of local parameters to de-noising, and then proposed a block matching algorithm which is called two-step strategy acceleration tactics, we found it can reduce the number of iterations and ensure the de-noising result.