RealText-V2-Syn25k

RealText-V2-Syn25k is a synthetic forgery dataset for document tampering detection, containing 25,028 high-quality forged document images generated from RealText-V2 authentic samples.

Dataset Description

We adopt a similarity-guided synthetic method [1] that generates forgeries from the clean images within the RealText-V2 samples. Specifically, the synthetic method uses two trained selection and quality models to automatically select source-target crop pairs from these clean images and produce high-quality forgeries across five manipulation types: copy-move, splicing, insertion, inpainting, and coverage.

Usage

The dataset is distributed as a compressed tar archive (syn.tgz). Extract it with:

tar -xzvf syn.tgz

Training Data Composition

The first 13,500 samples are sourced directly from RealText-V2 (6,000 authentic + 7,500 forged). The remaining 25,028 samples are synthetic forgeries generated from RealText-V2 clean images using the similarity-guided pipeline.

Manipulation Types

• Copy-move: A region is copied and pasted to another location in the same document.
• Splicing: A region from one document is transplanted into another document.
• Insertion: New text content is inserted into the document image.
• Inpainting: A region is removed and filled in with surrounding content.
• Coverage: A region is overlaid or covered with content from elsewhere.

References

[1] Mohamed Dhouib, Davide Buscaldi, Sonia Vanier, and Aymen Shabou. "Leveraging Contrastive Learning for a Similarity-Guided Tampered Document Data Generation Pipeline." arXiv preprint arXiv:2602.17322 (2026).

Source Dataset

RealText-V2 — a dataset of authentic and forged document images.