Research On Mixed Methods For Image Restoration

As one of the main ways for people to obtain and transmit information,the image is indispensable in daily life and scientific research.However,due to the existing technologies and equipments imperfect,there exists a certain degree of quality degradation,which may affect the subsequent image processing,in the process of image acquisition and transmission.Therefore,how to restore the high quality image has become the research direction of many scholars.Image denoising is an important branch of image restoration.In order to remove the noise better,researchers have proposed many methods.Aiming at image denoising problem,theoretical analysis and innovation experiments based on two existing denoising methods which have good performance are did in this thesis.Three effective denoising algorithms are proposed and proved by experiments.Because of its good denoising performance,the block matching and 3D filtering(BM3D)algorithm is often used to measure the effectiveness of a new method.However,BM3 D algorithm still has some limitations to protect the edges and details of the image.Therefore,this thesis propose an improved method based on directional diffusion equation to modify the BM3 D algorithm.Firstly,the BM3 D denoised image is decomposed into three subbands which are perpendicular to each other by the wavelet decomposition.Secondly,by using the directional diffusion equation which has the anisotropic diffusion operator and the diffusion coefficient,we diffuse the subband of the high-frequency component,of which the wavelet coefficients are less than the threshold of the BM3 D restored image to the corresponding subband of the noisy image.Finally,by replacing the corresponding subband of the BM3 D restored image with the new one we obtain the improved denoising result.Experimental results demonstrate that the proposed method is more effective than the BM3 D algorithm in reducing the noise and protecting the image edges and details.Similarly,the non-local mean image denoising algorithm is improved.The subband of the high-frequency component,of which the wavelet coefficients are less than the threshold of the NLM restored image contains many information of the original image.Therefore,we shrink the wavelet coefficients of its method noise image by using Bayes Shrink threshold and then add the result to the corresponding original subband of the NLM restored image.Experimental results show that the improved algorithm has better denoising performance than the NLM denoising algorithm in visual or objective.When processing the image with a high level of noise,the denoising performance of the BM3 D algorithm will decline and the restored image will appear fuzzy,artifacts and other issues.Therefore,this thesis put forward a two-stage image denoising method for high noise image based on the principal component analysis.First,the BM3 D algorithm is used to denoise the noisy image in the first stage.Then,for the method noise that the BM3 D creates,using the local pixel grouping method to select the training samples.After coefficient shrinkage in the PCA domain,we get the residual image by inverse PCA transform.Finally,adding the first stage BM3 D restored image and the residual image and,then using the BM3 D algorithm to denoise the new low noisy image to get the final result.Experimental results show our method can reduce the artifacts efficiently and preserve the edges better for high noise image compared to the original BM3 D algorithm.The corresponding PSNR and SSIM values are also improved.