Abstract: Highlights•This paper proposes LDH-Net, a dual-branch luminance-guided network that effectively addresses complex document image shadows while preserving text integrity.•This paper introduces Horizon-Vertical Attention and Dilated Convolution Mamba modules to capture both global dependencies and local details with low computational complexity.•This paper achieves state-of-the-art results across multiple benchmarks and demonstrates strong robustness and efficiency in real-world OCR applications.
External IDs:dblp:journals/ivc/YangLJWLW25
Loading