Efficient burst super-resolution network with effective feature processing and group up-sampling

Abstract

Burst super-resolution (Burst SR) has achieved significant performance improvements by utilizing burst images obtained from modern handheld devices. However, there are still several challenges related to Burst SR. Firstly, misalignment between burst images is caused by hand tremors. Secondly, aligned features should be effectively combined and up-sampled to harness the rich information of the burst images. Numerous recent studies have focused on performance enhancement by suggesting complex networks using a pre-trained optical flow network or transformer to address these problems. However, computational complexity is an essential consideration for limited-resource environments. In this paper, we present an Efficient Burst super-resolution Network (EBNet) with effective feature processing and group up-sampling strategies. For feature processing, we introduce novel denoising and refinement module to efficiently perform the alignment with deformable convolution instead of a pre-trained network. Furthermore, we propose the enhanced group up-sampling module to successfully merge burst features without loss of salient details. With these two strategies, our proposed EBNet remarkably alleviates the trade-off between performance and computational cost compared to the existing state-of-the-art.