.png){kind=icon}
.png){kind=icon}
Return
IAA-VSR: An iterative alignment algorithm for video super-resolution
.png){kind=icon}
PRE
AI
DOI：10.1007/s10489-022-03364-z.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Video Super-Resolution (VSR) aims at producing a high-resolution video from its corresponding low-resolution input frames. In VSR, the key to generating high-quality output is to exploit the spatial similarity of temporal frames. Most VSR methods achieve this by super-resolving a single reference frame with the aid of multiple frames in a temporal window. For this goal, some alignment methods have been proposed to compensate for the motion between adjacent frames. However, these methods lack more upper-level and unified guidance to progressively align neighboring frames, which often leads to poor results when encountering large motions. In this paper, we propose a unified Iterative Alignment Algorithm (IAA) for more accurate frame alignment in VSR. In IAA, each adjacent frame only needs to be aligned to its nearest neighbor, which greatly eases the alignment problem for all kinds of motions. To show the effectiveness of our method, we apply IAA to red the Enhanced Deformable Video super-Resolution (EDVR) network and obtain a new network called IAA-VSR. Extensive experiments show that our IAA-VSR consistently improves the performance of EDVR on benchmark datasets.
Keywords:
Video super-resolution
Image processing
Computer vision