Study On Image Denoising Based On New Multiscale Geometric Analysis

With the rapid development of the society,the ways of our communication are constantly changing.Image has become the most important means except the writing.While in the process of the image collection,acquisition and transmission,it will be subject to noise interference so that it can be resulted in unclear,twist and even distortion.So,processing is becoming more and more important,and image denoising plays a fundamental role.At present,image denoising methods based on machine learning,statical model and nonlocal means are the most popular methods which can obtain excellent results.But they still need to be improved.In this paper,image denoising algorithm based on new multiscale geometric analysis has been studied and main research works are as following:1.Image denoising method using the new HMT model based on non-subsample Contourlet conversion.With the help of Weibull to show the statistical properties of the coefficients after decomposition.Based on the coefficient of significance or the lack of meaning coefficient is the basis for judging whether the coefficient is significant or not.Taking the full advantage of the coefficient relativity and then Byaesian denoising is implemented on the probability obtained after training.The results show that better denoising not only protects the edge and texture,but also achieves the subjective expected results.2.A new denoising method is proposed based on the improved NLM filter with exponential moment in frequency domain.Weight value of Item similarity is calculated through using exponential moment to replace the traditional coefficient and using Modified Bisquare kernel function to substitute for the Gaussian kernel function.Results indicate that noise is delaminated and the marginal information is well reserved.3.Pixels that include required information are successfully picked out by TWSVM before the denoising using NLM method,which greatly reduces pixels need to be calculated.Meanwhile,the algorithm,Local Adapted James–Stein Type Center Pixel Weights is improved and the kernel functions are optimized by introducing the soft threshold.Denoising time is highly decreased and denoising effect is enhanced.