Deformable Aligned Fusion for Video Super Resolution

The key to the video super-resolution algorithm is to capture the information of adjacent frames to supplement the reconstruction of the current frame. It is necessary to apply the alignment operation to compensate the displacements of objects in different frames. We propose an innovative deformable aligned fusion block (DAFB) to exploit the complementary information from adjacent frames more efficiently. We also model the offsets of deformable alignments as probabilistic distributions to augment the robustness of a DAFB. Our proposed model outperforms state-of-the-art CNN-based models on benchmark testing datasets, including Vid4 and UDM10.

2023 Asia Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC)