Image super-resolution based on segmentation and classification with sparsity

Single image super-resolution reconstruction is a challenging ill-posed inverse problem currently, which aims to recover a high-resolution image from a low-resolution input. Natural images contain multiple types of textures or multiple object categories, and the features, especially the high-frequency feature, of each category are remarkably different from others. In this paper, we propose a method based on image segmentation and classification with sparse representation for single image super-resolution reconstruction. We divide the input low-resolution image into several regions, and find out the most correlative category for each segment. Then the high-resolution image for each segment is reconstructed by the sparse coefficients with respect to the dictionaries jointly learned from high-resolution and low-resolution image pairs of corresponding category respectively. The experimental results demonstrate the superior of our method over the several state-of-the-art methods both quantitatively and perceptually.