Image And Image Sequence Restoration Based On Regularization Method

The qualities of images and image sequences are degraded for some undesirable factors in the process of imaging and transmission including noise, interference, blur, low sampling rate and so on, which affect subsequent processing and application. It is paid more attentions in many fields that how to restore the degraded image and sequence into a perfect one with clear content removing the degeneration of quality. Image restoration is a process to improve the quality and visual effect of the degraded images, including the tasks of denoising, deblurring, super resolution reconstruction and so on, belonged to the category of the inverse problem in math.Based on the analysis of image degraded models, the thesis studies three problems including Poisson noisy image deblurring, image sequence denoising and deblurring, image super resolu-tion based on the regularization methods.(1) The technology of Poisson image deblurring employing mean curvature regularization is studied. Combining the mean curvature information of an image and the information divergence data fidelity term, the mean curvature regularization based Poisson image restoration model is proposed, which can maintain the edges and detail information, the mean curvature regularization can also preserve corners as well as image gray-scale intensity contrasts. Moreover, a numerical algorithm based on the augmented Lagrange multiplier method with a splitting technique is pro-posed to solve the model which is hard to handle for the high order and nonlinear characteristics. The proposed method can change the original hard problem into some subproblems which are easy to solve or with closed solutions. The proposed method can achieve good results for solv-ing the problem of un-blind Poisson image deblurring for natural image and biological image, Poisson image blind deblurring and image deblurring with multiplicative Gamma noise. The technology of Poisson image deblurring based on patch manifold regularization is also studied. Combining the patch manifold prior knowledge of an image, a regularization which can con-strain the smoothness of image is employed to address the Poisson image restoration problem. In order to improve the robustness, a robust parametric manifold learning method is proposed to present the distance between image and patch manifold, based on which the regularization and the model can be gotten. Furthermore, a numerical algorithm based on Gauss-Seidel iteration scheme and alternating direction method is proposed to solve the model. Experiments show that the proposed approach can produce higher quality results, especially for keeping the texture and details comparing with some state of the art methods in the field.(2)The technology of image sequence restoration adopting spatial-temporal mean curvature regularization is studied. The mean curvature of an image is expanded to a 3D image sequence data. The corresponding spatio-temporal mean curvature regularization is also proposed to en-hance the smoothness of the solution in the image domain and time domain. Combining the regularization and idea of space-time volume, the proposed approach treats an image sequence as a space-time volume both for the regularization and data fidelity term. The proposed method is employed to solve the video denoising with Gaussian noise, video deblurring with spatial invariant kernel and video deblurring with motion spatial variant kernel.(3)The technology of multiframe image super resolution using total generalized variation (TGV) regularization is studied. Considering the piecewise polynomial priori of an image, the TGV regularization is used to address the super resolution problem. In order to keep the detail of the image well and overcome the over-smoothed result from the non-adaptive parameter. The adaptive parameters are obtained by combining and adopting the local statistic characteristic of the image. In order to solve the non-smooth model, the half quadratic based optimization al-gorithm is used. After a smooth approximation to the original objective function by the Huber function, the equivalence problem can be solved by direction iterative algorithm. The experi-mental result shows that the proposed model and algorithm can well improve the resolution of the image and keep the information of texture and detail in the image.