Abstract
Photographs captured by smartphones and mid-range cameras have limitedspatial resolution and dynamic range, with noisy response in underexposedregions and color artefacts in saturated areas. This paper introduces the firstapproach (to the best of our knowledge) to the reconstruction ofhigh-resolution, high-dynamic range color images from raw photographic burstscaptured by a handheld camera with exposure bracketing. This method uses aphysically-accurate model of image formation to combine an iterativeoptimization algorithm for solving the corresponding inverse problem with alearned image representation for robust alignment and a learned natural imageprior. The proposed algorithm is fast, with low memory requirements compared tostate-of-the-art learning-based approaches to image restoration, and featuresthat are learned end to end from synthetic yet realistic data. Extensiveexperiments demonstrate its excellent performance with super-resolution factorsof up to $\times 4$ on real photographs taken in the wild with hand-heldcameras, and high robustness to low-light conditions, noise, camera shake, andmoderate object motion.