Research On Image3D-denoising Algorithm Based On Local Feature

As the important role of digital image in our production and life, and its beingvulnerable to noise pollution in the processing of generation and transmission whichaffects subsequent image processing, the effective image denoising in image processing isparticularly significant.After the study of basic knowledge and method of image denoising, this paper focuseson the promising denoising algorithms, the nonlocal means method and theblock-matching and3D collaborative filtering method. In order to further improve thedenoising performance and execution efficiency, the paper also puts forward two kinds ofimprovement: BM3D combining Tetrolet prefiltering and BM3D based on adaptivedistance hard threshold (BM3D-ADT). The work mainly includes:1. In order to change the bad denoising performance of BM3D in high noise level,BM3D denoising combining Tetrolet prefiltering is proposed. Firstly prefilter the imagewith the Tetrolet transform, then BM3D is used to denoise. Tetrolet’s well-preservation ofimage structure and part of noise being removed improve the accuracy of block-matchingin BM3D. Experiment results show that it can restrain noise better, and retain the imagerich details more precisely at the same time. It also greatly weakened the pseudo noisephenomenon.2. BM3D denoising based on adaptive distance threshold is put forward, whosedistance threshold is adaptively set according to the ratio of mean and standard deviationof the image and noise estimates. Firstly divide the input image into blocks. Thenaccording to gradient information, compute the structure similarity and Euclidean distancebetween reference block and all the other candidate blocks, and select correspondingEuclidean distance when the structure is the most similar. And then estimate the mean,standard deviation of each reference block, and the standard deviation of noise. With thedata fitting method, conclude the function relation between corresponding Euclideandistance and the ratio of mean-standard deviation of image and the standard deviation ofnoise, which is an adaptive distance threshold of BM3D. It has excellent denoisingperformance in terms of PSNR value and human visual perception, as well as reduces therunning time when noise level is low.