Tampered Text Recognition on TFR CTM Cross-Method
45.6OCR AccuracyTextShield-R1
Evaluation results
| Method | Links | |
|---|---|---|
| TextShield-R1Fine-tuning=Full training set images2026.02 | 45.6 | |
| InternVL3-8BFine-tuning=None2026.02 | 31.5 | |
| GPT4oFine-tuning=None2026.02 | 27.2 | |
| Qwen2.5-VL-7BFine-tuning=Full training set images2026.02 | 26.3 | |
| FakeShield*Fine-tuning=Full training set images, Base MLLM=Qwen2.5-VL-7B2026.02 | 26.3 | |
| SIDA*Fine-tuning=Full training set images, Base MLLM=Qwen2.5-VL-7B2026.02 | 26.3 | |
| InternVL3-8BFine-tuning=Full training set images2026.02 | 23.6 | |
| Qwen2.5-VL-7BFine-tuning=None2026.02 | 21 | |
| MiniCPM_V_2.6Fine-tuning=Full training set images2026.02 | 20.7 | |
| Qwen2.5-VL-3BFine-tuning=Full training set images2026.02 | 18.8 | |
| InternVL3-2BFine-tuning=Full training set images2026.02 | 18.7 | |
| InternVL3-2BFine-tuning=None2026.02 | 18.2 | |
| FakeShieldFine-tuning=Full training set images2026.02 | 15.1 | |
| SIDAFine-tuning=Full training set images2026.02 | 15.1 | |
| Qwen2.5-VL-3BFine-tuning=None2026.02 | 7.7 | |
| MiniCPM_V_2.6Fine-tuning=None2026.02 | 7 |