EBSR: Feature Enhanced Burst Super-Resolution With Deformable Alignment

Ziwei Luo, Lei Yu, Xuan Mo, Youwei Li, Lanpeng Jia, Haoqiang Fan, Jian Sun, Shuaicheng Liu; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2021, pp. 471-478

We propose a novel architecture to handle the problem of multi-frame super-resolution (MFSR). The proposed framework is known as Enhanced Burst Super-Resolution (EBSR), which divides the MFSR problem into three parts: alignment, fusion, and reconstruction. We propose a Feature Enhanced Pyramid Cascading and Deformable convolution (FEPCD) module to align multiple low-resolution burst images in the feature level. And then the aligned features are fused by a Cross Non-Local Fusion (CNLF) module. Finally, the SR image is reconstructed by the Long Range Concatenation Network (LRCN). In addition, we build a cascading residual pathway structure (CR) to improve the performance. We conduct several experiments to analyze and demonstrate these modules. Our EBSR model won the champion in the real track and second place in the synthetic track in the NTIRE21 Burst Super-Resolution Challenge.