Recent developments in the field of imaging sensors reveal a trend towards high frame rates while pixel resolution seems to converge towards a physical limitation. Usage of mobile devices like smartphones spreads rapidly, often including primitive but high-framerate video functionality. Working in the field of telemedicine for infrastructurally underdeveloped regions, my current research considers improving image quality of simple imaging devices to allow for in-field telemedical assistance. Imaging processors of current mobile devices use excessive noise reduction to improve visual appeal of the raw data coming from tiny, noisy sensors. This leads to smearing of low contrast parts of the image thus awkwardly rendering important details like human skin.
Therefore I have developed two algorithms for enhancing image quality of mobile devices through combining several frames to one. This technique is already known as super-resolution. The goal is to improve resolution (and noise) characteristic of images by exploiting its detail coherence to predecessor and successor frames.
The first algorithm is very fast and qualifies for processing of live streams. It is self-adaptive to noise and the number of frames combined, producing very naturally looking image quality, certainly much better than any single source frame.
The second algorithm takes lots of processing time thus qualifying only for rendering of stills. It is also self-adaptive to noise and detail level and is scalable for 1-5x enlargement.
Currently, I am working on publications of this algorithms and will present some experimental results here in my blog soon. For now it seems that my machine learning based algorithms perform at least as good as current commercial implementations like photo acute.