Improved example-based single-image super-resolution

How to reduce the super-resolving time for realtime application but not to tamper observably with the quality of image is the interesting pivotal point of research about example-based single-image super-resolution now. The paper proposes a method which classifies these high-frequency patches of low-resolution image to accelerate the procedure. Before the super-resolution, classify method is used to mark the high-frequency patches of the low-resolution image with corresponding labels. During superresolution, the distances between each matching patches of low-resolution image and middle-frequency patches of training set are computed. The candidate patch is the patch within training set with the minimum distance. For the patches labeled with non-edge, few candidates are selected, while for flat patches the matching step can be canceled, directly replacing high-resolution patches by enlarged interpolation of low-resolution patches. Two examples are use to illustrate the performance of the proposed algorithm, one using a factitious image obtained by blurring and down-sampling an original image, and another using directly a true image. The results show the proposed method can reduce effectively the computational complexity.