Image Super-Resolution via Sparse Representation and Local Texture Constraint

Recent researches have proved the efficiency of sparse representation based methods in single image super resolution (SISR) reconstruction. However, these classic methods fail to consider the image edge information and local texture details during reconstruction process, thus result in unwanted edge artifacts and local texture blurring of the reconstructed high-resolution (HR) images. In this paper, considering that the HR image and its corresponding low-resolution (LR) image share similar texture structure in corresponding positions, we propose a new SISR reconstruction method by combining sparse representation and a local texture constraint. In our method, the HR and LR image patch pairs are firstly extracted from training samples and then are used to train a HR and LR dictionary pair. Then, in code stage, we apply a local texture constraint to restrict the local texture similarity between the input LR image patches and reconstructed HR image patches. Furthermore, we introduce a global texture constraint to a global optimization model to further enhance the reconstructed image quality. Experimental results prove that the proposed method can generate sharper edges and clearer texture details than some state-of-the-art image super-resolution methods.