Divide and Conquer: Rethinking HDR Reconstruction utilizing HSV color space

The reconstruction of High Dynamic Range (HDR) images from multiple exposures of Low Dynamic Range (LDR) images presents significant challenges, especially in the presence of camera shake and dynamic object motion. Despite the promising results achieved by existing deep learning methods, they often entail substantial training cost. Utilizing the HSV color space informed by the physiological attributes of human vision, this paper proposes a novel lightweight dual-branch HDR reconstruction network (LightHDR) framework. Specifically, we employ the Value (V) channels of differently exposed images for deghosting, and utilize the Hue (H) and Saturation (S) channels of the reference image for chromatic reconstruction. These two branches operate independently and in parallel, ultimately reconstructing an HDR image. Experiments on benchmark dataset demonstrate that our network achieves an optimal equilibrium between HDR reconstruction quality and computational cost, outperforming the state-of-the-art HDR algorithms.