Document Shadow Removal Guided by Joint Text-Prior and Shadow Probability

Current document shadow removal methods mainly focus on estimating shadow-free background colors but overlook foreground text extraction. This often causes artifacts and text fragmentation in the resulting images. To tackle these drawbacks, we propose a novel approach guided by foreground text and shadow probability. By incorporating shadow probability map estimation and text mask estimation modules, our method supports targeted shadow elimination while retaining text regions, which substantially boosts image restoration performance. We decompose a shadowed document image into foreground text and shadowed background via filtering operations. First, we predict text masks for the foreground to protect text information throughout shadow removal. Next, shadow probability is estimated from the shadowed background. Finally, document shadows are removed under the joint guidance of shadow probability maps and text masks. The three task-specific networks for text mask estimation, shadow probability estimation, and shadow removal are all built upon Attention U-Net as the backbone to produce high-quality restored images. Experimental results show that the proposed method outperforms existing methods in terms of both quantitative metrics and visual quality. It can effectively eliminate shadow artifacts while maintaining the integrity of text content.