Research For Image Super-resolution Reconstruction Algorithm Based On Sparse Representation

Image super-resolution is an important image processing technique, which can improve the visual effects of images and prepare for the post-processing of images. Over the last decade, with the development of machine learning and pattern recognization, learning-based image super-resolution algorithms are widely studied and have achieved better reconstruction results than classic methods. Image super-resolution methods based on sparse representation are one kind of them, which have been developed in recent years and achieved state-of-the-art results. This paper has mainly studied the application of sparse representation in image super-resolution from two respects: dictionary training methods in sparse representation and the methods of establishing sparse regularization terms. In the final chapter, this paper proposed that applying the multiple image super-resolution method based on sparse representation to spatio-temporal data fusion of remote sensing images.In traditional image super-resolution methods based on MAP, the noise model is usually based on Gaussian or Laplacian distribution. However, it is difficult for the noise model based on one of the two distributions to deal with adaptively complex noise in real scenes. To solve this problem, this paper proposed a new hybrid noise model, which adaptively decides the noise distribution trend via estimating the distribution parameters of noise.This paper classified the image super-resolution algorithms into two classes according to the difference of processed low resolution images: ones based on single image and ones based on multiple images. Further, this paper introduced two kinds of effective dictionary training methods: dictionary pair based on image features and dictionary based on structure classification; at the same time, this paper summarized the methods of establishing regularization terms according to different sparse constraints: ones based on local sparse constraint, ones based on local weighting sparse constraints and ones based on global sparse constraint. Finally, this paper analyzed and compared the features of these methods through experiments.In another respect, since remote sensing sensors can not acquire images having both high spatial resolution and high temporal resolution, this paper proposed applying sparse representation methods to this problem. Considering features of remote sensing sequence, this paper developed two kinds of methods in available image feature domain: ones based on spatial information of remote sensing images and ones based on temporal information of remote sensing images. Finally, these two methods are compared with the classic methods of spatio-temporal image fusion via experiments.