Learning interlaced sparse Sinkhorn matching network for video super-resolution

作者：

Highlights：

• An efficient and effective ISSM module has been designed for feature alignment. To the best of our knowledge, this is the first work to learn optimal matching for VSR.

• A Bi-TF model has been developed for temporal fusion, which demonstrates much better performance than the commonly-used fusion techniques in VSR.

• Extensive evaluations on three benchmark datasets have demonstrated competing performance of the proposed method against the state-of-the-art methods.

摘要

•An efficient and effective ISSM module has been designed for feature alignment. To the best of our knowledge, this is the first work to learn optimal matching for VSR.•A Bi-TF model has been developed for temporal fusion, which demonstrates much better performance than the commonly-used fusion techniques in VSR.•Extensive evaluations on three benchmark datasets have demonstrated competing performance of the proposed method against the state-of-the-art methods.

论文关键词：Video super-resolution,Multi-scale feature,Interlaced sparse sinkhorn attention,Bidirectional fusion,Dynamic reconstruction

论文评审过程：Received 1 April 2021, Revised 25 August 2021, Accepted 29 November 2021, Available online 6 December 2021, Version of Record 24 December 2021.

论文官网地址：https://doi.org/10.1016/j.patcog.2021.108475