Learning approaches have shown great success in the task of super-resolvingan image given a low resolution input. Video super-resolution aims forexploiting additionally the information from multiple images. Typically, theimages are related via optical flow and consecutive image warping. In thispaper, we provide an end-to-end video super-resolution network that, incontrast to previous works, includes the estimation of optical flow in theoverall network architecture. We analyze the usage of optical flow for videosuper-resolution and find that common off-the-shelf image warping does notallow video super-resolution to benefit much from optical flow. We ratherpropose an operation for motion compensation that performs warping from low tohigh resolution directly. We show that with this network configuration, videosuper-resolution can benefit from optical flow and we obtain state-of-the-artresults on the popular test sets. We also show that the processing of wholeimages rather than independent patches is responsible for a large increase inaccuracy.
展开▼
