基于稀疏表示的单幅图像超分辨率-信号与信息处理专业论文.docx

I II Abstract Abstract This paper researches the field of image super-resolution, namely using single or a group of low-resolution images to reconstruct a high-resolution image. As for the ill-posed nature of multi-image super-resolution, the paper focuses on single-image super-resolution, studies the mainstream approaches in the field and performs a series of improvements for the sparse representation based algorithm. The primary work of this paper can be summarized as follow: Firstly, in terms of the algorithm model, we describe the image degradation model and the principle of algorithm based on sparse representation. Two image super-resolution algorithm of that category are detailed, which are respectively to analyze advantages and disadvantages. Then we put forward our own improvements. High- and low-resolution image blocks are simultaneously delt with to train joint dictionary. What’ s more, the principle of iterative back-projection is utilized to strengthen global constraints. Secondly, in terms of numerical solution in the process of sparse solving, the pixel value of high-resolution image block is too large which hinders operational speed and causes arithmetic error. Taken into account the IBP algorithm proposed Irani et al., which converges fast and rebuilds good results, we calculate the difference images between IBP’s results and high-resolution images to train high-resolution dictionary. At last, non-local similarity redundance of image is extracted to establish a post-processing procedure to correct the reconstructed image. In the end, in terms of the adopted dictionary, analytically designed dictionaries lack adaptability, while the reconstructed images from a universal and over-complete dictionary tend to generate undesirable artifacts. Drawn from the adaptive sparse domain selection procedure proposed by Dong et al., we preserve the PCA sub-dictionaries clustered by K-means but update image blocks with dictionary through an iterative process.