Research On Structure Self-similarity Based Image Denoising

As an important information carrier, image is widely used in many fields. Due to the limitation of imaging system and imaging environment, it's inevitable to suffer from noise and blurs which greatly restrict the subsequent high level visual tasks, such as segmentation, detection and recognition. The difficulty of image denoising lies in that it's difficult to accurately model the noise characteristic and to balance the noise suppression and detail preservation. This paper conducts the image denoising research based on structure self-similarity, including the analysis and modeling for both noise and image characteristic, and the fast numerical algorithm. Our dissertation mainly includes the following aspects:First, this paper introduces the image structrure self-similarity and corresponding modeling method. The key of image denoising depends on finding and ultilizing the structure self-similarity. The local smoothness, non-local similarity and cross-scale similarity of the image are introduced. The tight framelets with multi-scale and multi-direction ability to describe the sparsity of the image is introduced. And low-rank approach is used to model the non-local structure self-similarity. Which makes firm foundation for the modeling of self-similarity based denoising methods.Secondly, aiming at the complex noise property of SAR, double data fidelity terms are proposed to accurately model the mixed noise. By adjusting the weights of two data fidelity terms type unknown noise can be well modeled. Furtherly incorporate the non-local self-similarity prior by using low-rank approach, the SAR despeckling model is builded up. And this method can effectively deal with despeckling.Further more, denoising method combining local and non-local self-similarity for space target images with property of “bright target and dark background” is proposed. By the fusion of sparsity under framelet domain and the low-rank characteristic in non-local similar patches, the solution space is well regularized. ADMM method can be used to iteratively solve this model. Except the good noise suppression, the details can also be preserved.Finally, in order to take the advantage of spectral smoothness of hyperspectral images, a spectral-spatial total variation based image restoration method is proposed. The hyperspectral images can be regarded as a three dimensional data cube, and thus spatial and spectral smoothness priors can be used simultaneously to maintain the spectral information of hyperspectral data cube during denoising.