Super resolution via sparse representation in l1 framework

This paper proposes a new high speed Single Image Super Resolution algorithm and also suggests modifications that can perform super resolution on video sequences. Embarking from recent successful algorithms proposed by Yang et. al.[18] and Elad et. al.[19], it adds a number of enhancements that improve both PSNR of the recovered image and performance of the dictionary training. It also proposes an incremental dictionary update strategy that enhances results on video sequences by improving the dictionary quality at each frame. The algorithm does not necessarily need a training image set, though it can use one to enhance PSNR of the upscaled image. It automatically picks the patches that would benefit from super resolution, ignoring others, thus enhancing speed. It also partially accounts for spatial transformations of patches in the trained dictionary, further enhancing performance