LDH-Net: Luminance-based Deep Hybrid Network for Document Image De-shadowing

Abstract

Existing deep learning-based Document Image De-shadowing (DID) methods face two key challenges. First, they struggle with complex shadows due to insufficient use of auxiliary information, such as shadow locations and illumination details. Second, they fail to effectively balance global relationships across the entire image with local feature learning to restore texture details in shadowed regions. To address these limitations, we propose a dual-branch de-shadowing network, called LDH-Net, which integrates luminance information as an auxiliary information for de-shadowing. The first branch extracts shadow-distorted features by estimating a shadow luminance map, while the second branch uses them to locate shadow regions and guide the de-shadowing. Both branches employ a hybrid feature learning mechanism to capture local and global information efficiently with lower complexity. This mechanism includes two key modules: Horizon-Vertical Attention (HVA) and Dilated Convolution Mamba (DCM). HVA models long-range pixel dependencies to propagate contextual information across the entire image to ensure global coherence and consistency. DCM utilizes dilated convolution within the State Space Model (SSM) to capture extensive contextual information and preserve local image details. Additionally, we introduce a luminance map loss to provide accurate optimization for reconstruction. Experiments on RDD, Kligler’s, Jung’s, and OSR demonstrate that LDH-Net outperforms previous state-of-the-art methods. Specifically, LDH-Net achieves the best PSNR/SSIM/LPIPS scores across all datasets, with up to 37.76 PSNR/0.981 SSIM/0.005 LPIPS on RDD datasets and consistent improvements on other benchmarks, confirming its superior performance on both visual quality and structural preservation.