Superresolution Videos and Optical Flow based on Combinatorial and Variational Optimization

点击翻译按钮获取中文摘要

Humans have the capability to draw very precise information from video even in case of very bad image quality. This astonishing capability becomes evident only when we look at the single images of a video, where we find out how noisy and blurry they typically are. This is also true in the era of HD videos, which formally have a high resolution, yet due to natural limitations in recording, single frames cannot provide the same quality as photos with the same resolution. In this project, the information of successive frames of a video are sought to be combined in a way that the quality and resolution of all frames can be increased. As the central hypothesis we claim that optical flow estimation, denoising, and superresolution are coupled problems. Thus, based on extensive prior work in these areas, we will develop techniques that simultaneously compute very precise optical flow and denoised single frames at a higher resolution. Concrete subprojects are concerned with the modeling of motion blur, fast motion, and occlusions in the context of video superresolution. We believe that joint optimization of optical flow and superresolution will provide new opportunities in video analysis, it will enable the restoration of old movies, and it will allow to lift existing low resolution videos to modern HD resolution (and beyond).