Research map / Image generation / Text-to-image diffusion
Revisiting Tampered Scene Text Detection in the Era of Generative AI
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.
Introduces open-set tampered scene text detection with a new OSTF benchmark (8 text editing methods), a Texture Jitter pre-training paradigm, and the Difference-Aware Forensics (DAF) pluggable branch on Faster/Cascade R-CNN.
Problem
Existing tampered scene text detectors trained on one forgery type fail to generalize to unseen forgery types produced by newer generative text editing models, and the Tampered-IC13 benchmark only covers SRNet-based forgeries.
Approach
Pre-train by subtly altering the texture of randomly chosen text regions (blur, sharpen, JPEG compression) and training the detector to localize these altered regions, plus a DAF branch that classifies text by its feature difference from a learned authentic kernel rather than from raw features.
Key results
- On Tampered-IC13 the zero-shot Ours+Cascade R-CNN reaches 81.7 mIoU versus 71.7 for full-shot UPOCR, and full-shot Ours+Faster R-CNN reaches 86.2 mIoU (segmentation comparison).
- On the OSTF ablation, Texture Jitter plus DAF with Cascade R-CNN raises tampered-text mF from 58.56/60.34 baseline to about 75.85 average mF over 81 test settings.
- Full-shot Ours+Cascade R-CNN reaches 88.62 mean F-score on Tampered-IC13 versus 76.66 for the prior SOTA (S3R+CounterNet, 76.7).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Texture Jitter pre-training instead of text detection pre-training (ablation) | Faster R-CNN · OSTF | tampered-text mean F1-score: 45.78 -> 71.96 | Helped |
| use Texture Jitter pre-training instead of COCO detection pre-training (ablation) | Faster R-CNN · OSTF | tampered-text mean F1-score: 34.82 -> 71.96 | Helped |
| add DAF to Texture Jitter pre-training (ablation) | Faster R-CNN · OSTF | mean F1-score: 73.41 -> 74.82 | Helped |
| add SCL to Texture Jitter pre-training (ablation) | Faster R-CNN · OSTF | mean F1-score: 73.41 -> 73.79 | Helped |
| replace Faster R-CNN with Cascade R-CNN (ablation) | Texture Jitter + DAF · OSTF | mean F1-score: 74.82 -> 75.85 | Helped |
| add SRM noise features to RGB (ablation) | OSTF model · OSTF | mean F1-score: 74.8 -> 74.8 | No clear effect |
| add Bayar noise features to RGB (ablation) | OSTF model · OSTF | mean F1-score: 74.8 -> 73.8 | Hurt |
| add NoisePrint++ noise features to RGB (ablation) | OSTF model · OSTF | mean F1-score: 74.8 -> 74.4 | Hurt |
| replace DAF with Contrast loss (ablation) | OSTF model · OSTF | mean F1-score: 74.8 -> 73.9 | Hurt |
| replace DAF with Triplet loss (ablation) | OSTF model · OSTF | mean F1-score: 74.8 -> 73.6 | Hurt |
| replace DAF with SCL (ablation) | OSTF model · OSTF | mean F1-score: 74.8 -> 73.8 | Hurt |
| use a perfect text detector instead of the learned detector (ablation) | Ours + Faster R-CNN · Tampered-IC13 | mean F-score: 87.46 -> 97.98 | Helped |
Surprising findings
- Pre-training on purely generated text-editing outputs (MOSTEL, DST, AnyText) without manual refinement caused NaN losses and failure to converge.
- Adding single-center loss alone to Texture Jitter gave only tiny gains, while DAF gave much larger improvements.
Limitations
- Open-set performance is weaker on STEFANN (binary-like textures) and TextOCR (tiny fuzzy text).
- Benchmark forgeries are manually refined with PhotoShop/GIMP, which limits scale.
- Pipeline relies on a text detector; with a perfect detector performance rises substantially, indicating detector dependence.
Experimental setup
Datasets: Tampered-IC13 OSTF LSVT ReCTS ICDAR2013 ICDAR2015 ICDAR2017 TextOCR ArT COCO
Benchmarks: Tampered-IC13 OSTF
Metrics: pixel-level IoU pixel-level F1 instance-level precision instance-level recall instance-level F1 mean F1
Baselines: DeepLabV3+ HRNetV2 BEiT-UPer SegFormer Swin-UPer UPOCR S3R+EAST S3R+PSENet S3R+ATRR S3R+CounterNet DTD UniFD
Builds on: Faster R-CNN Cascade R-CNN S3R UPOCR single-center loss
Closest papers by meaning
- TextShield-R1: Reinforced Reasoning for Tampered Text Detection (AAAI 2026 · similarity 0.90)
- Toward Real-world Text Image Forgery Localization: Structured and Interpretable Data Synthesis (NeurIPS 2025 · similarity 0.89)
- FakeShield: Explainable Image Forgery Detection and Localization via Multi-modal Large Language Models (ICLR 2025 · similarity 0.89)
- Forensics-Bench: A Comprehensive Forgery Detection Benchmark Suite for Large Vision Language Models (CVPR 2025 · similarity 0.87)
- Towards General Visual-Linguistic Face Forgery Detection (CVPR 2025 · similarity 0.87)
- A Hitchhiker's Guide to Fine-Grained Face Forgery Detection Using Common Sense Reasoning (NeurIPS 2024 · similarity 0.86)
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.