Deformable Spatial-Temporal Attention for Lightweight Video Super-Resolution

Video super-resolution (VSR) aims to recover high-resolution video frames from their corresponding low-resolution video frames and their adjacent consecutive frames. Although some progress has been made, most existing methods typically use the spatial-temporal information of two adjacent reference frames to aid in enhancing the video frame super-resolution reconstruction effect. This makes it impossible for these methods to achieve satisfactory results. To solve this problem. We propose a deformable spatial-temporal attention (DSTA) module for video super-resolution. The deformable spatial-temporal attention module improves the reconstruction effect by aggregating favorable spatial-temporal information from multiple reference frames into the current frame. To speed up the model training, we select only the first s highly relevant feature points as the attention scheme. Experimental results show that our method with fewer network parameters has strong video super-resolution performance.