Research On Super-Resolution Image Reconstruction

Super-Resolution image Reconstruction (SRR) is the process in which allows the recovery of a high-resolution (HR) image from a sequence of low-resolution (LR) images that are noisy, blurred, and down sampled.Firstly, this thesis reviews the research development in the field of the Super-Resolution image Reconstruction. Based on the fundamental principle and relevant issues of SRR research, it was generalized into two kinds of image observation model.According to the characteristic of the image periodic boundary condition, we can transform the image observation model and divide the SRR process into two sub-processes, measurements fusing and filter estimation de-blurring, which could effectively release the burden of the program and the computation cost, without losing the quality of the reconstructed result.On the research of image registration, which formed the starting points of SRR research, D.Keren's registration method by using Gaussian pyramid structure is presented. Based on the distribution of optical filed, an alternative approach is explored, which could obtain the sub-pixel movement estimation through utilizing region statistic.Based on the image observation model and image registration, it explored the applications of the Moving Surface Fitting (MSF), adaptive filtering - Least Mean Square (LMS) estimation, Maximum a Posteriori (MAP) frame model, image prior model and regularization in SRR. In the course of research, in order to solve the problem that exits in traditional MAP frame algorithm, which often takes the Gaussian model as its image priori model that could lead to the detail loss of reconstructed image, this paper selects Huber-Markov random field (HMRF) as the image priori model, and proposes a new priori model that takes the Directional HMRF (DHMRF) as the smooth constraint condition, incorporated into the regularization item, on one hand, which can assure the astringency and stability of the solution to the ill-problem, the other preserve the edge information of the object in the image as much as possible. Experiment results confirm the proposed approach, which can compensate the un-directional HMRF, is not only robust to noise and model error, but also can better retain the details and dramatically improve the quality of the result image.