Deeply feature fused video super-resolution network using temporal grouping

The video super-resolution (VSR) task refers to the use of corresponding low-resolution frames and multiple neighboring frames to generate high-resolution (HR) frames. An important step in VSR is to fuse the features of the reference frame with the features of the supporting frame. Existing VSR methods do not take full advantage of the information provided by distant neighboring frames and usually fuse the information in a one-stage manner. In this paper, we propose a deep fusion video super-resolution network based on temporal grouping. We divide the input sequence into groups according to different frame rates to provide more accurate supplementary information. Our method aggregates temporal-spatial information at different fusion stages. Firstly, we group the input sequence. Then the temporal-spatial information is extracted and fused hierarchically, and these groups are used to recover the information lost in the reference frame. Secondly, integrate information within each group to generate group-wise features, and then perform multi-stage fusion. The information of the reference frame is fully utilized, resulting in a better recovery of video details. Finally, the upsampling module is used to generate HR frames. We conduct a comprehensive comparative experiment on Vid4, SPMC-11 and Vimeo-90K-T datasets. The results show that the proposed method achieves good performance compared with state-of-the-art methods.