Burst Super-Resolution with Adaptive Feature Refinement and Enhanced Group Up-sampling

Abstract

Burst super-resolution (Burst SR) has achieved significant performance improvements through burst images captured by modern handheld devices. However, Burst SR still encounters several challenges. 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. As numerous studies have attempted to solve these problems, the complexity of alignment and fusion strategies has been increased and the significance of up-sampling has been overlooked. It exacerbates the trade-off between performance and computational cost and leads to a problem where information from multiple frames is not effectively exploited in the up-sampling process. In this paper, we present a Burst SR network with adaptive feature refinement and enhanced group up-sampling strategies. To feature refine adaptively, we introduce a novel denoising and correction module to elaborately perform the alignment with deformable convolution instead of a pre-trained network. Furthermore, we propose the enhanced group up-sampling module to merge burst features without the loss of salient details effectively. With these two strategies, our proposed method achieves state-of-the-art performance and alleviates the trade-off between performance and computational cost compared to the existing Burst SR approaches.