Image Denoising Algorithm And Application Based On Multiscale Geometric Analysis And Partial Differential Equations

Image acquisition system has been widely used in various areas, receives more and moreattention. Because of the system and environmental factors, images acquired have a lot of noise whichaffects the image visual quality. Image denoising is a fundamental field in image processing. Thefollowing works become difficult because of these noisy images, so the image denoising isparticularly important in the whole image process system. The purpose of image denoising is toestimate the original image from noisy image. Recently, as the development in Wavelet technologyand multiscale geometric theory, image denoising becomes a extremely active topic in imageprocessing,pattern recognition and computer vision. Previous algorithms are lack of universalapplicability, so they are difficult to be applied in image process system.This paper discussed multiscale geometric theory and partial differential equations, presented animproved image denoising method; combined the theory of grey system and used grey relation theoryin coefficients selection and proposed an image denoising algorithm in multiscale geometric analysisdomain. To solve the drawbacks of traditional image quality assessment methods, No-Referenceimage quality assessment was built to assess the denoised images, and the results accord with humanvisual character. The main results of research are as follows:(1) A common model of partial differential equations-Total Variation model was analyzed indetail. To solve the drawbacks of iteration rule in Total Variation model, an improved model based onenergy balance was proposed. Improved model gave two kinds of stop rules and their applications. Itnot only reduced the computational complexity, but also achieved good de-noising result.(2) According to the situation of noise estimation in many image denoising algorithms, a newnoise estimation method was proposed. The Wavelet coefficients in each sub-band can be wellmodelized by a Generalized Gaussian Distribution (GGD). As a result, the parameters of GGD modelcan be used to estimate the noise variance. The new noise estimate method has the character ofself-adaption and obtains high precision.(3) Based on the analysis of partial differential equations and multiscale geometric theory, anovel image denoising algorithm using Total Variation model and grey relation theory in Waveletdomain was proposed. The new algorithm analyzed the Wavelet coefficients in different areas, anddecomposed image into two parts: low frequency area and high frequency area. In low frequency area, there is little noise information, so the improved Total Variation model was used to reduce the noiseand maintain the edge. In high frequency area, Wavelet coefficients were selected by using greyrelational system. This method proposed a better threshold procedure, it considered the coefficientsrelationship both in scale and location, and obtained the superior denoising result and less edgeoscillation than traditional methods.(4) To overcome the deficiency of directional information and shift variation in Wavelet domain,the application in the Nonsubsampled Contourlet domain was considered as a candidate. Based on theimage denoising algorithm using Total Variation model and grey relation theory in Wavelet domain,and combined with the traditional Nonsubsampled Contourlet domain threshold method, four modelswere proposed and the best one was used in the new method. Compared with the Non Local Means(NL-means) algorithm, the new method using Total Variation model and grey relation theory inNonsubsampled Contourlet domain has a very good performance and preserves most importantinformation.(5) In practice, Full-Reference methods may not be applicable because the image for reference isnot often available. As a result, a new No-Reference image quality assessment method based on anatural image statistic model in Wavelet transform domain was proposed. A Generalized GaussianDistribution model was used to modelize the marginal distribution of Wavelet coefficients, so that theparameters of GGD were used to evaluate the image quality. The method is easy to implement andefficient in computation. Furthermore, it can be applied to many well-known types of imagedistortions, and achieves good performance.(6) The new algorithms built in this dissertation are of important value in image denoising, alongwith the corresponding software, which provide an important tool for image denoising. Thesimulation software applied in the image acquisition system obtained a good visual sense and ensuredthe following works.