Research map / Detection and segmentation / CNN features
Enhancing Tampered Text Detection through Frequency Feature Fusion and Decomposition
Zhongxi Chen, Shen Chen, Taiping Yao, Ke Sun, Shouhong Ding, Xianming Lin, Liujuan Cao, Rongrong Ji
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add DCT Frequency Perception Head to the RGB segmenter (ablation) | ConvNextV2-Uper · DocTamper-FCD | IoU: 0.80 -> 0.83 | Improved |
| add the Visual Enhancement Module with zero-initialized conv fusion on top of FPH (ablation) | FFDN · DocTamper-T | IoU: 0.86 -> 0.88 | Improved |
| combine VEM and WFE together on top of FPH (ablation) | FFDN · DocTamper-T | IoU: 0.88 -> 0.90 | Improved |
| zero-initialized conv for frequency features instead of random init (ablation) | VEM · DocTamper-T | IoU: 0.882 -> 0.895 | Improved |
| use addition instead of max for fusing frequency and RGB features (ablation) | VEM · DocTamper-T | IoU: 0.881 -> 0.895 | Improved |
| use proposed CA+SA enhancement instead of CBAM (ablation) | VEM · DocTamper-T | IoU: 0.877 -> 0.895 | Improved |
| add attention enhancement to wavelet components (learnable) instead of fixed DWT (ablation) | WFE · DocTamper-T | IoU: 0.867 -> 0.895 | Improved |
| add high-frequency enhancement in FFM (vs low-frequency only) (ablation) | WFE · DocTamper-T | IoU: 0.874 -> 0.895 | Improved |
| add low-frequency enhancement in FFM on top of high-frequency (ablation) | WFE · DocTamper-T | IoU: 0.892 -> 0.895 | Improved |
| apply the FFDN modules to a Swin-S backbone (ablation) | Swin-Uper · DocTamper-T | F1: 0.73 -> 0.82 | Improved |
| use FFDN instead of DTD under JPEG Q75 compression | DocTamper-T · DocTamper-T, Q75 | IoU: 0.75 -> 0.82 | Improved |
| FFDN versus RGB-only ConvNextV2 on average drop at Q75 | DocTamper-T · DocTamper-T, Q75 | avg drop from Q100 (%): 17.8% -> 20.2% | Hurt |
Closest papers by meaning
- Revisiting Tampered Scene Text Detection in the Era of Generative AI (AAAI 2025 · similarity 0.86)
- Learning Discriminative Noise Guidance for Image Forgery Detection and Localization (AAAI 2024 · similarity 0.83)
- Toward Real-world Text Image Forgery Localization: Structured and Interpretable Data Synthesis (NeurIPS 2025 · similarity 0.83)
- TextShield-R1: Reinforced Reasoning for Tampered Text Detection (AAAI 2026 · similarity 0.83)
- FakeShield: Explainable Image Forgery Detection and Localization via Multi-modal Large Language Models (ICLR 2025 · similarity 0.82)
- Towards Better Robustness Against Natural Corruptions in Document Tampering Localization (AAAI 2025 · similarity 0.82)
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.
