Abstract: Existing document image shadow removal methods neglect soft shadow characteristics, causing residual shadow edges and color distortion. A novel algorithm is proposed based on shadow density extraction and guidance. First, precise shadow density maps are generated in the document image shadow removal dataset through a designed shadow density calculation model to independently train shadow density extraction network (SDENet). Subsequently, with SDENet frozen, shadow density fusion module (SDFM) fuses shad-ow density features and document shadow image features to guide the separate training process of the document shadow removal network (DSRNet) for initial shadow removal. Finally, joint training of SDENet and DSRNet with parameter fine-tuning achieves improved performance. Experimental results demonstrate that compared to state-of-the-art methods, the proposed algorithm improves SSIM and PSNR by 0.73% and 6.25% respectively while reducing RMSE by 21.00% on the RDD dataset, and improves SSIM and PSNR by 0.16% and 11.35% respectively while reducing RMSE by 36.22% on the SD7K dataset.