Research map / Image generation / Text-to-image diffusion
Toward Real-world Text Image Forgery Localization: Structured and Interpretable Data Synthesis
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes FSTS, a Fourier-series-inspired hierarchical model of real-world text-image tampering parameters collected from 67 human editors, and uses sampled configurations to synthesize training data that improves generalization of text image forgery localization models.
Problem
Text image forgery localization (T-IFL) models generalize poorly to real-world tampering because scarce real datasets and synthetic datasets have a distribution gap in the invisible tampering-parameter distribution.
Approach
Collect real tampering operation parameters from human editing traces (video, PSD, logs), model them hierarchically as weighted combinations of representative basis operation-parameter configurations at individual and population levels (Fourier-series analogy), and sample from this distribution to synthesize training images.
Key results
- Models trained on FSTS-T outperform DocT-T on real-world datasets in Protocol 2, with average F1 gains above 14% across methods (e.g., TruFor average F1 0.198 -> 0.477 for DocT-T vs FSTS-T on the real sets average row).
- Under Protocol 4, FSTS-T pretraining followed by CertificatePS fine-tuning improves average F1 for most methods, e.g., TruFor average F1 0.459 direct vs 0.498 with FSTS-T pretraining.
- FSTS collects 16,750 real tampering instances from 67 participants; the released FSTS dataset totals 294,182 images.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train on FSTS-T instead of DocT-T | TruFor · Real-world average (Protocol 2) | F1: .198 -> .477 | Helped |
| Train on FSTS-T instead of DocT-T | RRU-Net · Real-world average (Protocol 2) | F1: .199 -> .342 | Helped |
| Train on FSTS-T instead of DocT-T | DFCN · Real-world average (Protocol 2) | F1: .102 -> .327 | Helped |
| Train on FSTS-T instead of DocT-T | PSCC-Net · Real-world average (Protocol 2) | F1: .227 -> .389 | Helped |
| Train on FSTS-T instead of DocT-T | MVSS-Net · Real-world average (Protocol 2) | F1: .168 -> .386 | Helped |
| Train on FSTS-T instead of DocT-T | DTD · Real-world average (Protocol 2) | F1: .097 -> .345 | Helped |
| Train on FSTS-T instead of DocT-T | STFL-Net · Real-world average (Protocol 2) | F1: .205 -> .399 | Helped |
| Train on FSTS-T rather than DocT-T before fine-tuning on CertificatePS | TruFor · Real-world average (Protocol 4) | F1: .453 -> .498 | Helped |
| Train on FSTS-T rather than DocT-T before fine-tuning on CertificatePS | RRU-Net · Real-world average (Protocol 4) | F1: .324 -> .415 | Helped |
| Train on FSTS-T rather than DocT-T before fine-tuning on CertificatePS | DFCN · Real-world average (Protocol 4) | F1: .307 -> .400 | Helped |
| Train on FSTS-T rather than DocT-T before fine-tuning on CertificatePS | PSCC-Net · Real-world average (Protocol 4) | F1: .416 -> .418 | No clear effect |
| Train on FSTS-T rather than DocT-T before fine-tuning on CertificatePS | MVSS-Net · Real-world average (Protocol 4) | F1: .406 -> .434 | Helped |
Surprising findings
- Models trained on DocT-T often score worse than the direct real-data baseline on cross-dataset real tests, while FSTS-T pretraining helps most methods.
- Several synthetic-trained models (e.g., PSCC-Net on AFAC) remain weak on specific real datasets despite the overall gains.
Limitations
- Post-processing frequencies scaled by 0.3 by heuristic choice to avoid overly complex samples.
- Modeling is approximate: assumes representative configurations t_k match real-world ones, and reduces distribution matching to coefficient alignment.
- Some methods still perform poorly on specific real datasets (e.g., PSCC-Net on AFAC, DTD on STFD).
Experimental setup
Datasets: DocTamper CertificatePS AFAC STFD FindIt FSTS-T FSTS-S FSTS-1.5k TIC13 T-SROIE OSTF
Benchmarks: DocTamper-Test FSTS-S FSTS-1.5k AFAC CertificatePS STFD FindIt
Metrics: pixel-level F1 AUC
Baselines: RRU-Net DFCN PSCC-Net MVSS-Net TruFor DTD STFL-Net DocTamper
Builds on: DocTamper PS-scripted
Closest papers by meaning
- FakeShield: Explainable Image Forgery Detection and Localization via Multi-modal Large Language Models (ICLR 2025 · similarity 0.91)
- Revisiting Tampered Scene Text Detection in the Era of Generative AI (AAAI 2025 · similarity 0.89)
- TextShield-R1: Reinforced Reasoning for Tampered Text Detection (AAAI 2026 · similarity 0.89)
- SUMI-IFL: An Information-Theoretic Framework for Image Forgery Localization with Sufficiency and Minimality Constraints (AAAI 2025 · similarity 0.88)
- ForensicHub: A Unified Benchmark & Codebase for All-Domain Fake Image Detection and Localization (NeurIPS 2025 · similarity 0.87)
- Forensics-Bench: A Comprehensive Forgery Detection Benchmark Suite for Large Vision Language Models (CVPR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.