Multi-Prior Based Multi-Scale Condition Network for Single-Image HDR Reconstruction

High Dynamic Range (HDR) imaging aims to reconstruct the natural appearance of real-world scenes by expanding the bit depth of captured images. However, due to the imaging pipeline of off-the-shelf cameras, information loss in over-exposed areas and noise in under-exposed areas pose significant challenges for single-image HDR imaging. As a result, the key to success lies in restoring over-exposed regions and denoising under-exposed regions. In this paper, a multi-prior based multi-scale condition network is proposed to address this issue. (1) Three types of prior knowledge modulate the intermediate features in the reconstruction network from different perspectives, resulting in improved modulation effects. (2) Multi-scale fusion extracts and integrates deep semantic information from various priors. Experiments on the NTIRE HDR challenge dataset demonstrate that the proposed method achieves state-of-the-art quantitative results.