Super-resolution reconstruction based on structure tensor's eigenvalue and classification dictionary

Abstract

Super-resolution provides effective prior information for the single-frame super resolution reconstruction. It's difficult to recover fine grained details via a general dictionary, trained through the diversified training samples, due to the negli-gence of structural characteristics. Thus, the dictionary whic-h is adaptive to local structures is needed. Considering the eigenvalues of structure tensor are convenient to distinguish the edge regions and the smooth regions, we construct a sca-lar texture feature descriptor to present texture information of image. It is employed for clustering low and high resolution patches in the training stage and for model selection in the reconstruction stage. Tight sub-dictionary is learned for each cluster. For a given test image patch, the corresponding sub-dictionary is adaptively selected, and then super-resolution reconstruction of this image is completed. Com-pared with the recently proposed dictionary learning met-hods for image super-resolution reconstruction, the algorit-hm preserves more details and ensures the quality of the reconstructed image.