Single-Image HDR Reconstruction by Multi-Exposure Generation
Winter Conference on Applications of Computer Vision (WACV) 2023 · Oct 11, 2022
High dynamic range (HDR) imaging is indispensable in modern photography, but traditional multi-image pipelines struggle with alignment, ghosting, and fusion artifacts. Single-image HDR reconstruction offers a promising alternative: learn to map pixel values directly to irradiance, bypassing the align-and-merge pipeline entirely.
This work proposes a weakly supervised method that inverts the physical image-formation process — the network learns to generate multiple exposures from a single image, reconstructing pixel irradiance and hallucinating detail in under- and over-exposed regions. Training requires only pairs of under-/over-exposed images (no HDR ground truth), using a representation loss, a reconstruction loss, and a perceptual loss. The method achieves state-of-the-art results on the DrTMO dataset.