Watermark removal is essential for Optical Character Recognition and digital document reconstruction. Document images differ significantly from natural scene images, necessitating tailored image processing methods. Traditional Transformer-based methods, though effective, face scalability issues due to the quadratic complexity of attention computations. Inspired by Mamba, which scales linearly with context length, we introduce a semantically-guided Mamba approach, MambaDW. This two-stage method first removes most watermarks and then employs semantic-guided 2D selective scanning for precise enhancement and global information fusion. We also incorporate a self-supervised loss to enhance generalizability on unlabeled real data. Our experiments show MambaDW’s effectiveness in document watermark removal.