Super Resolution: A Database Driven Inference Approach
Monica A. Trifas, Jeremy Straub · 2011
A revised super-resolution technique is presented and evaluated in this paper. This technique incorporates prior knowledge gained via training subject domain specific or general purpose images prior to presenting an image for super-resolution image resolution enhancement. When presented with an image to enhance, the engine selects candidate patterns using a lowresolution search mechanism and then uses the higher sourceimage resolution to select a winning candidate for inclusion in the super-resolved image. This proposed technique is evaluated from an application-agnostic perspective using several evaluation metrics. The failure of common super-resolution evaluation metrics to adequately measure application goal success is discussed. This failure is also directly tied to engine performance as, in many cases, the success metric and the pattern selection metric are tightly aligned. Thus, by developing new super-resolution evaluation metrics and other enhancements, not only can evaluation of training-based superresolution be improved but also the super-resolved image product as well.