A Rough Note on Example Based Super Resolution

This note is a rough summary of the super resolution methods that I recently learned. To begin with, what’s the definition of super resolution? Super Resolution (SR) is a class of techniques that enhance the resolution of an image. For example, given a 256 × 256 image as input, the output of a SR process is a corresponding 512× 512 image. Notice that super resolution is a highly ill-posed problem. Because for a given low resolution (LR) image, there exists infinitely many (a large number at the same scale, infinitely many scales) high resolution (HR) images that can produce the same LR image. Therefore, the goal of super resolution is to impose different kinds of priors to narrow down the space of HR images that when downsampling, produce an indentical LR image as the input. Roughly speaking, the techniques of super resolution can be categorized into three class: