Study On Single Image Super-Resolution

Image super-resolution reconstruction algorithm is an active research topic in computer vision.It is an algorithm for obtaining high resolution images from one or more low resolution images.However,due to the ill-posedness of image super-resolution reconstruction,we need to utilize some suitable prior knowledge to constrain the solution.At present,many algorithms can achieve satisfactory results,but some still may be affected by outliers,and produce many artifacts.In this thesis,we propose two algorithms for super-resolution reconstruction:We propose an SR reconstruction method for single image based on non-local sparse and low-rank regularization,inspired by the non-local redundancy of natural images,which can be used to recover the latent features and high frequency details.We group a set of similar patches for each extracted patch and form them as a matrix.This matrix can be decomposed as low rank component and sparse component,where the low rank component comes from the similarity and the sparse component comes from the fine differences and outliers.Considering these two constraints and combining with the fidelity term,we propose our algorithm model.Using the property that l_1/2,norm is sparser than l1 norm,we propose an image super-resolution algorithm based on l1/2 regularization,non-local sparsity and low rank regularization.We use l1/2 norm instead of l1 norm to constrain the sparse representation coefficient.Then,combining the coupled dictionary training and sparse representation algorithm,we can obtain the preliminary result of image super-resolution reconstruction.In order to further optimize the preliminary result,we apply the iterative back projection theory to project the results into the solution space of degradation model.Moreover,during the iteration process,we use the non-local sparse and low rank regularization to constraint the result.We qualitatively and quantitatively compare our algorithm with other state-of-the-art algorithms,and the results show that our proposed methods are superior to others.