Bhat Goutam ; Danelljan Martin ; Van Gool Luc ; Timofte Radu

Abstract

While single-image super-resolution (SISR) has attracted substantial interestin recent years, the proposed approaches are limited to learning image priorsin order to add high frequency details. In contrast, multi-framesuper-resolution (MFSR) offers the possibility of reconstructing rich detailsby combining signal information from multiple shifted images. This keyadvantage, along with the increasing popularity of burst photography, have madeMFSR an important problem for real-world applications. We propose a novel architecture for the burst super-resolution task. Ournetwork takes multiple noisy RAW images as input, and generates a denoised,super-resolved RGB image as output. This is achieved by explicitly aligningdeep embeddings of the input frames using pixel-wise optical flow. Theinformation from all frames are then adaptively merged using an attention-basedfusion module. In order to enable training and evaluation on real-world data,we additionally introduce the BurstSR dataset, consisting of smartphone burstsand high-resolution DSLR ground-truth. We perform comprehensive experimentalanalysis, demonstrating the effectiveness of the proposed architecture.