Stamp removal from real documents: A diffusion-based approach with EA-RMSE evaluation

Stamps, which are vulnerable to forgery and often interfere with digital document processing, are increasingly recognized as elements that need to be removed to ensure the reliability and usability of digital documents. Existing stamp removal models predominantly rely on fully synthetic datasets, in which stamp patterns are digitally overlaid onto clean document images, and therefore fail to capture real physical effects such as print quality variation, stamp degradation, and scanning artifacts. Moreover, conventional image quality metrics like PSNR and SSIM fail to adequately reflect localized errors critical to document readability and usability. To address these limitations, we construct a real-world stamp dataset by collecting physically stamped documents under diverse conditions and propose a hybrid training strategy that integrates real-world and synthetic data. To further examine document-layout and language diversity, we complement the analysis with virtual-document real-stamp data, where multilingual and multi-type document backgrounds are generated while stamps are physically applied. Additionally, we introduce Error-Area Root Mean Square Error (EA-RMSE), a novel evaluation metric that separately quantifies residual stamps and missing document content to provide a more perceptually aligned assessment of stamp removal quality. Experimental results demonstrate that our method significantly improves robustness across datasets and provides more interpretable evaluations, offering a strong foundation for future research in stamp removal tasks.