Abstract:Existing shadow removal methods for document images often rely on strictly paired data or manually annotated masks, which tend to cause overfitting and poor generalization in real-world scenarios. To address this issue, this study proposes a background-aware semi-supervised shadow removal method for document images. The method adopts a two-stage framework that combines supervised and unsupervised learning. In the supervised stage, an image decomposition unit is employed to separate the input into a shadow-free reflectance map and a shadow-containing illumination map. The reflectance map is used to constrain unsupervised training to maintain structural consistency, while the illumination map highlights brightness differences in shadow regions to reduce learning difficulty. Moreover, a background color estimation module and a shadow attention map are introduced to provide global color guidance and regional constraints. An offset estimation module is also incorporated to correct color and illumination deviations, enhancing structural recovery and shadow modeling capabilities. In the unsupervised stage, guided by the decomposition unit and trained on unpaired data, the deshadowing network is optimized using adversarial loss, color-brightness loss, and reflectance consistency loss to improve robustness and generalization. Experimental results show that the proposed method achieves superior performance on document image shadow datasets and produces more natural and refined shadow-free results in real document scenes.