Super-resolution Image Reconstruction Based On Mutual Information Registration

Super-resolution image reconstruction is a technique that merges the complementary information and redundant information retained in defferent images about the same scene, base on the relative motion relationship of these images. Super-resolution technique use multiple frames of images to reconstruct a high resolution image, by interpolation, deblur and denoise operations. It is a signal processing method that gains the spatial resolution from time efficiency. Super-resolution can solve the problem that a image obtained through hardware cannot meet the need of high-resolution in same case such as high cost, poor hardware condition and hard imaging environment. In this paper, we aim to establish a system which can reconstruct image details effectively by using super-resolution signal processing technology.In the article, we first analyse the math and physical principle of Super-resolution technique. Then we study the imaging process, build a imaging model and establish a equation to describe the relationship between low resolution observed images and the high resolution imaging target. Based on the equation, we make a deeper analysis of the Super-resolution problem. The Super-resolution process consists of two important process, that is, image registration and image reconstruction, which is the research point in the article.Image registration refers to the determination process of relative motion relationship between different observed images. A high precision image registration is the basis of accurate image reconstruction. We mainly study the mutual-information registration method, which has been widely used, especially in multi-modal registration. Considering the drawbacks of the maximum mutual-information, we add edge information to mutual-information criterion, through dual-tree complex wavelet transformation(DT-CWT). At the same time, we also make an improvement of partial volume(PV) interpolation method in the registration process, to decrease the probability of falling into the local point. All the mutual-information registration process is based on the DT-CWT pyramid modal.As to the image reconstruction work, we first propose a interpolation method based on dual-tree complex wavelet transformation, which is applied in the two image reconstruction methods proposed in the following content. The two reconstruction methods are the image merging method based on DT-CWT interpolation and DT-CWT merging, a improved POCS(projecting on to convex sets) method. Referring to the improvement of the POCS, we choose a much better initial image in the reconstruction process; propose a adative loose operator to accelerate the speed of POCS method; propose a parallel projecting structure in case of no convergence in the POCS algorithm; energy non-decreasing constraint and image smoothness constraint are also applied, to make sure the the image quality and overcom the edge shocking effect.Based on the aboved study, this article achieves the goal of a complete image resolution reconstruction process. The experiments prove that the mutual-information registration method and the reconstruction method proposed in the article are effective.