Deep Unrolled Weighted Low-Rank Approximation for High Dynamic Range Imaging

The rapid advancement of digital imaging technology has led to significant improvements in image quality. However, conventional sensors often struggle to capture the full dynamic range of natural scenes, leading to under-and over-exposed regions in images. To address this challenge, high dynamic range (HDR) imaging techniques have been developed, with one effective approach involving the synthesis of an HDR image from multiple low dynamic range (LDR) images captured at varying exposures. Despite the progress made in HDR imaging, issues such as ghosting artifacts caused by motion between exposures remain a challenge. In this paper, we introduce an interpretable deep unrolling approach for HDR imaging that leverages the strengths of both model-based and learning-based methods. Our approach employs Weighted Low-Rank Approximation (WLRA) to guide the reconstruction process of the missing information in the irradiance matrix and introduces two regularization functions learned through deep unrolling to address potential inaccuracies. Experimental results show that our proposed algorithm achieves significantly higher scores in PU-PSNR, PU-MSSSIM, and HDR-VDP compared to conventional methods, demonstrating its effectiveness in producing high-quality, ghost-free HDR images with minimal dependence on large training datasets. This work provides a robust solution that bridges the gap between theoretical models and deep learning, offering greater interpretability and generalization in HDR imaging.