Luo Ziwei ; Li Youwei ; Cheng Shen ; Yu Lei ; Wu Qi ; Wen Zhihong ; Fan Haoqiang ; Sun Jian ; Liu Shuaicheng

Abstract

This work addresses the Burst Super-Resolution (BurstSR) task using a newarchitecture, which requires restoring a high-quality image from a sequence ofnoisy, misaligned, and low-resolution RAW bursts. To overcome the challenges inBurstSR, we propose a Burst Super-Resolution Transformer (BSRT), which cansignificantly improve the capability of extracting inter-frame information andreconstruction. To achieve this goal, we propose a Pyramid Flow-GuidedDeformable Convolution Network (Pyramid FG-DCN) and incorporate SwinTransformer Blocks and Groups as our main backbone. More specifically, wecombine optical flows and deformable convolutions, hence our BSRT can handlemisalignment and aggregate the potential texture information in multi-framesmore efficiently. In addition, our Transformer-based structure can capturelong-range dependency to further improve the performance. The evaluation onboth synthetic and real-world tracks demonstrates that our approach achieves anew state-of-the-art in BurstSR task. Further, our BSRT wins the championshipin the NTIRE2022 Burst Super-Resolution Challenge.