Super-resolution with probabilistic motion estimation

Classic super-resolution has long relied on very exact motion estimation for the recovery of sub-pixel details. As a highly accurate motion field is hard to obtain for general scenes, classic super-resolution has been known to be limited to specific cases, where the motion is of a global nature. In this chapter, we present a recently developed family of algorithms that shatters this barrier. These novel algorithms relax the requirement of a one-to-one motion field, and replace it with a simple, probabilistic motion estimation. The probabilistic motion field is integrated into the classic (and heavily investigated) SR framework, and ultimately results in a very simple family of algorithms. The obtained paradigm gets an algorithmic structure that resembles that of the nonlocal means, and as such, leads to a localized and easily parallelizable procedure. Despite their simplicity, the obtained algorithms are nevertheless very powerful in handling the most general scenes, with the probabilistic motion estimation enabling the handling of challenging motion patterns. The resulting image sequences are of high quality, and contain few artifacts. These novel algorithms open the door to a new era in super-resolution that bypasses the limiting traditional reliance on explicit motion estimation for super-resolution.