DDSR-Net: Direct Document Shadow Removal Leveraging Multi-Scale Attention

Shadows in document images are undesirable yet inevitable. They can decrease the clarity and readability of the images. The existing methods for removing shadows from documents still face some challenges, such as the traditional heuristics lack universality and the optimization goal of subnetworks is not consistent for multistage deep learning methods. In this paper, we introduce an end-to-end direct document shadow removal network (DDSR-Net), where we employ a 3-layer Unet++ as the backbone to extract features from diverse scales. To further improve the performance of DDSR-Net, we integrate the multiscale attention (MSA) blocks into each node. The MSA block allocates different weights to feature vectors at different positions, achieving automatic feature alignment and significantly enhancing the end-to-end network’s ability to handle shadow processing. To verify the effectiveness of the proposed DDSR-Net, qualitative and quantitative experiments are conducted on multiple open-source document shadow removal datasets. The experimental results demonstrate that our method outperforms the existing state-of-the-art methods on these datasets. Our code and models will be released to the public.