A Study On Image Super-resolution Reconstruction

Vision sense is the main way for human to acquire information. Images, as a true reflection of the objective world, becomes a information carrier that human com-monly use to understand the world. In our lives, The facts of imaging device hardware cost, manufacturing process and transfer limitations make the acquired images be low-resolution. These low-resolution images restrict us to exploit them further. This casts an image super-resolution reconstruction problem, which aims to reconstruct a latent clear high-resolution image from one or more degraded low-resolution images while requires the restored image to be as clear as possible, to be as natural as possible, and contain as little artifacts as possible.In this thesis, we regard the super-resolution reconstruction as the inverse of im-age degradation process. So we can solve this problem by denoising, up-sampling, deblurring processes. Hence, we propose a multi-point filter based on steering kernel regression, as well as a novel image blind deconvolution approach based on difference evolution algorithm.The proposed multi-point filter combines the advantages of steering kernel regres-sion method and the guided filter method. Steering kernel regression method can effec-tively estimate the filter kernel under high-level noise condition. We employ the filter kernel to weight the analysis window for guided filter, to make the local line model as-sumption more reasonable. Although guided filter has outstanding edge-preserve prop-erty, it has common denoising performance. The designed multi-point filter inherits the edge-preserve property of guided filter, meanwhile improve the guided filter’s denois-ing performance. Compared with the steering kernel regression method, our methods can produce denoising results which approximate to the iterative results of complex regression model. When comparing with the state-of-the-art methods, our multi-point filter has comparable denoising performance.We present a novel blind deconvolution approach based on difference evolution method fully utilizes the simple processes of its optimization to optimize the complex image blind deconvolution problem. In order to overcome the defects of naive Bayesian inference, we design a delicate prior on blur kernel to avoid no-blur explanation. Con-sidering the blur kernel has fewer variables than image, we choose the blur kernel as the input of difference evolution algorithm. Due to the iteration nature of evolution algorithm, the objective cost should be calculated many times. In order to quicken computation speed, we relax the image prior to Gaussian prior, and then use the sparse prior for final clear image estimation. The experiments validate our method, and our method can succeed in processing defocus blurring and motion blurring.In order to improve the quality of restored image further, this thesis focuses on studying image blind super-resolution reconstruction problem, i.e. estimating the latent high-resolution image and meanwhile estimating the blur kernel in degradation process. First, we analyze the facts which can affect the accuracy of estimated super-resolution blur kernel, then present a super-resolution kernel estimation algorithm based on useful edges under Bayesian inference framework. As a result, the accurate estimated super-resolution blur kernel make the results more sharper. It is worthy to mention that al-though the image super-resolution problem has been cast long ago, there is no algorithm which can deal with any images. It is because that any algorithm has its own assumed degradation model. When the observed image satisfy the model, it will produce nice results. However, when the observed image deviates from assumed model, the restored image will have low-quality. The experiments on synthetic images validate our method, and our method can improve the performance of other algorithms without blur estima-tion.In the end, this thesis first cast an image fitting multiple displays problem, this problem contains image retargeting and super-resolution reconstruction problems si-multaneously. Hence, we propose a content-aware image blind super-resolution method to up-sample input image to any resolution, and meanwhile ensure the reconstructed im-age with sharp edges and that the salient regions can not be distorted. In addition, we for-mulate a novel sparse matrix, derived from the retargeting method, has the similar func-tion to down-sampling matrix in super-resolution problem. We call it as content-aware down-sampling matrix. We replace the down-sampling matrix in super-resolution prob-lem with the novel matrix. Then, we can solve image retargeting and super-resolution reconstruction problem simultaneously. In terms of quantity and quality, our method is effective and efficient.