SAFHDR: Single-exposure HDR reconstruction via attention-guided feature learning and global information aggregation
Lucas Hildelbrano Costa Carvalho, Quefren Oliveira Leher, Josue Lopez-Cabrejos, Cristopher Ochoa-Villanueva, Gustavo de Souza Ferreti, Karine Firmo, Thuanne Paixão, Facundo Palomino-Quispe, Ana Beatriz Alvarez, Olacir Rodrigues Castro Junior · Computers & Graphics · 2026
The reconstruction of High Dynamic Range (HDR) images from a single Low Dynamic Range (LDR) image is a challenging problem in computer vision. State-of-the-art methods still have limitations in recovering regions degraded by severe clipping and noise in overexposed and underexposed areas. To overcome these limitations, this paper proposes SAFHDR, a U-Net-based architecture composed of DAMU-Net, which incorporates deformable convolutions guided by the Optimized Attention Block and efficient multiscale processing via the Lightweight Residual Features Block, and the Information Aggregation Module, dedicated to preserving the global structure of the scene. Additionally, a two-stage training strategy is adopted to improve perceptual fidelity and suppress chromatic and structural artifacts. Experiments conducted on the NTIRE 2021 dataset demonstrate that the proposed method outperforms state-of-the-art methods on HDR-oriented perceptual metrics, with high detail recovery capability in critical regions.