Image Super-resolution Based On Regularization Model

With the development of display device, e.g., high-quality video,high-quality projector, high-quality camera, etc., the previous images orvideos cannot meet with the requirement of these hardware. Besides, theresolution of medical or surveillance image/video is highly required, sothat the afterward analysis can be done better. The aim of super-resolutionis to make these images or videos successfully displayed by the hardware.Specifically, super-resolution can increase the resolution of input image orvideo, making them attain high-resolution or even super-resolution visualresults. This thesis focuses on the three main technologies insuper-resolution, i.e., interpolation-based method, reconstruction-basedmethod and example-based method. We investigate the existing problemsof these three technologies, and propose our solutions. We mainly researchon how to apply the regularization model of scale-invariance property ontothe image degradation model.For the interpolation-based method, the traditional ways are able tohandle only one intrinsic dimension, while natural images are composed ofthree kinds of intrinsic dimensions. For tackling all dimensions, wepropose to combine directional interpolation and bi-cubic interpolation.This method is able to deal with both edges and checkerboard-like patternsin natural images.For the reconstruction-based method, the traditional methods enhancethe gradient field via pixel level, which is of little efficiency. According tothis, we propose a patch-based gradient enhancement algorithm, which ishighly effective and can achieve robust super-resolution results. Besides, we propose three solutions to the optimization problem of imagedegradation model with gradient regularization. The experimental resultsshow that our proposed method is able to present lower computationcomplexity and better subjective evaluation.For the example-based method, we mainly focus on two works. First,we propose a quaternion-based sparse representation for imagesuper-resolution. Instead of only considering single channel forsuper-resolution, our proposed method takes all three color channels intoconsideration, thus preserve colors better. Second, we propose a methodthat combines self-similarity and fractal-based invariance property.Equipped with phase congruency analysis, our proposed super-resolutionis able to adaptively up-sample the edge, texture and flat regions, whichachieves even better visual results.We choose some common images in image processing as testingdataset, and apply the magnification factor of4for up-sampling. Wecompare our proposed method with most state-of-art methods. Theexperimental results show that our proposed method not only achieveshigher objective evaluation, but also get better subjective evaluation.