Publikation
DocForgeNet: Dual Cross-Stream Fusion Network for Robust Forgery Detection in Scanned Documents
Nauman Riaz; Stefan Agne; Andreas Dengel; Sheraz Ahmed
In: International Conference on Document Analysis and Recognition. International Conference on Document Analysis and Recognition (ICDAR-2025), The 19th International Conference on Document Analysis and Recognition, located at ICDAR-2025, September 16-21, Hubei, Wuhan, China, Springer Nature, 2025.
Zusammenfassung
Document image forgeries, especially text manipulations in scanned documents, pose severe risks to information integrity, impacting domains such as legal, financial, and personal records. These alterations are often subtle, localized, and visually indistinguishable, especially under common compression standards like JPEG. Existing detection methods typically rely on single-model architectures, either convolutional neural networks (CNNs) or transformers, which independently struggle with effectively capturing both local artifacts and global structural inconsistencies. To overcome these limitations, we propose DocForgeNet, a novel dual cross-stream fusion network explicitly designed for robust detection and localization of forged text regions in document images. DocForgeNet simultaneously processes RGB and discrete cosine transform (DCT) features through parallel CNN and transformer streams. The CNN stream excels at identifying local inconsistencies, such as compression artifacts and font irregularities, while the transformer stream leverages self-attention mechanisms to model broader, contextual discrepancies indicative of large-scale manipulations. Cross-linear attention modules facilitate effective feature fusion between the streams without information loss. Extensive evaluations on the DocTamper dataset, which contains forgeries on 170,000 document images of various types, demonstrate that DocForgeNet significantly outperforms state-of-the-art methods, achieving superior precision, recall, and F1-scores across various testing scenarios, including severe JPEG compression. Our approach not only establishes a new benchmark in tampering detection performance but also highlights the effectiveness of integrating complementary local and global representations for enhanced document integrity verification.