Research Of Image Super-resolution Reconstruction Based On Sparse Representation Theory

Image super-resolution（SR） reconstruction is to reconstruct ahigh-resolution（HR）image from one or a series of low-resolution（LR）images in the same scene with a certain amount of prior knowledge. Suchtechnique can effectively enhance image resolution, improve the visualeffect and prepare a good foundation for the later processing of the image,without changing the existing equipments.This dissertation first reviews some classical SR algorithms whichbelong to the interpolation based methods, the reconstruction based waysand the learning based schemes, respectively. Among these algorithms,we discuss the SR scheme based on sparse representation proposed byYang in detail. Compared with the traditional SR algorithms, the Yang’sscheme performs significantly better in the recovery results. But it needsa large amount of test images during the dictionary training stage, and itcost a long time for the HR image restoring. To solve the abovementioned problems, this paper separately proposed two algorithms, aMCA decomposition based SR algorithm and a saliency map based SRalgorithm. The MCA decomposition based SR algorithm utilizes MCA todecompose an image into the texture part and the structure part and onlytakes the texture part to train the dictionary. The reconstruction of thetexture part is based on sparse representation, while that of the structurepart is based on more faster method, the bicubic interpolation. As forsaliency map based algorithm, it decomposes an image into the salientpart and the non-salient part, according to the vision character of humaneyes. The new approach only applies sparse representation based SR algorithm on salient area, and uses the bicubic interpolation to reconstructthe non-salient area.Experiments show that the proposed methods perform better eitherin the dictionary training process or at the time-costing aspect, whilewithout reducing the quality of the recovered HR images.