Image Super-Resolution Using Trillions of Examples

University Of Virginia Main Campus, Charlottesville VA

Investigators

Abstract

Many important image processing tasks involve solving some type of inverse problem. Examples include: eliminating noise, compensating for low quality optical systems, enhancing a black-and-white image with plausible colors and improving the resolution of an image. All are ill-conditioned and so their solutions must incorporate assumptions about natural images. Although researchers have made astounding progress on these problems in the last fifty years, a key limiting property of current techniques is that they analyze their input more or less in isolation. This research will ask the question of whether an image can be improved by evaluating trillions of image patches constructed from millions of on-line images and using the set of relevant patches to improve the quality of the original. This work will focus on one particular inverse problem: super-resolution, or increasing the resolution of an image to reveal missing details. The IBM/Google compute cluster in conjunction with the MapReduce programming framework will be instrumental in developing and evaluating these data-intensive algorithms. This research will investigate scaling existing example-based techniques to use a massive training database and develop entirely new techniques that better capitalize on this amount of data by incorporating higher-level patterns in images such as scene categories and object boundaries.