Reconstructing HDR Image from a Single Exposure LDR Image with Global Feature Fusion

Modern digital cameras face inherent limitations when capturing the full luminance range in real-world scenarios, primarily due to hardware constraints. Additionally, challenging lighting conditions, such as severe over-exposure, often result in image detail distortion. Most CNN-based HDR methods are further hindered by their limited receptive fields, impeding their ability to recover fine details in overexposed areas. To address these issues, this paper introduces a novel architecture featuring a UNet-like base branch, complemented by a global feature fusion module. This module seamlessly blends local and global features, facilitating multi-scale information extraction for image reconstruction. Furthermore, the architecture incorporates a weighted branch to preserve well-exposed details and a global branch for capturing global features. By adopting this approach, our method not only achieves state-of-the-art performance in quantitative comparisons but also obtains excellent visual quality in assessments. In our experiment, we achieved a PSNR-μ score of 34.04 and an HDRVDP3 score of 8.311, outperforming the previous state-of-the-art methods by improvements of 0.02 and 0.01, respectively.