Research map / Detection and segmentation / CNN features
CatmullRom Splines-Based Regression for Image Forgery Localization
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 CSR-Net, which recasts image forgery localization as regression by fitting CatmullRom splines to tampered-region contours, and adds a comprehensive re-scoring step and a vertical/horizontal texture module to suppress false positives and sharpen boundaries.
Problem
Image forgery localization suffers from false positives from threshold-sensitive binarization in segmentation methods and inaccurate, irregular tampered-region boundaries.
Approach
Converts mask annotations into CatmullRom spline control points so that anchor-free regression can be applied to a pixel-level task; scores each region by combining classification and instance scores, and uses directional (1xk and kx1) learnable edge heatmaps to refine contours.
Key results
- Pre-trained CSR-Net reaches 94.3% AUC on Coverage, 88.1% on CASIA, 88.3% on NIST16 and 85.4% on IMD20, best among compared methods, but 96.8% on Columbia trails PSCCNet by 1.4 points.
- After fine-tuning, CSR-Net achieves 97.9 AUC / 78.0 F1 on Coverage, 90.4 AUC / 58.5 F1 on CASIA and 99.7 AUC / 83.5 F1 on NIST16.
- Ablation on CASIA shows removing the CRA module drops AUC from 90.4 to 78.1 (w/o CSR baseline) and removing VTP lowers F1 from 58.8 to 56.9.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add CatmullRom splines instead of traditional regression (ablation) | CSR-Net · CASIA | AUC: 68.5 -> 90.4 | Helped |
| Add CatmullRom splines instead of traditional regression (ablation) | CSR-Net · NIST16 | AUC: 75.9 -> 99.7 | Helped |
| Add Comprehensive Re-scoring Algorithm (CRA) (ablation) | CSR-Net · CASIA | AUC: 86.3 -> 90.4 | Helped |
| Add Comprehensive Re-scoring Algorithm (CRA) (ablation) | CSR-Net · CASIA | F1: 55.5 -> 58.8 | Helped |
| Add Comprehensive Re-scoring Algorithm (CRA) (ablation) | CSR-Net · NIST16 | AUC: 95.9 -> 99.7 | Helped |
| Add Comprehensive Re-scoring Algorithm (CRA) (ablation) | CSR-Net · NIST16 | F1: 78.8 -> 83.5 | Helped |
| Add Vertical Texture-interactive Perception (VTP) (ablation) | CSR-Net · CASIA | AUC: 88.9 -> 90.4 | Helped |
| Add Vertical Texture-interactive Perception (VTP) (ablation) | CSR-Net · CASIA | F1: 56.9 -> 58.8 | Helped |
| Add Vertical Texture-interactive Perception (VTP) (ablation) | CSR-Net · NIST16 | AUC: 97.8 -> 99.7 | Helped |
| Add Vertical Texture-interactive Perception (VTP) (ablation) | CSR-Net · NIST16 | F1: 81.8 -> 83.5 | Helped |
| Use CatmullRom splines rather than the traditional regression baseline (ablation) | CSR-Net · CASIA | F1: 35.3 -> 58.8 | Helped |
| Use CatmullRom splines rather than the traditional regression baseline (ablation) | CSR-Net · NIST16 | F1: 51.2 -> 83.5 | Helped |
Surprising findings
- Conventional CatmullRom tension (tau=1) fits poorly; tau=16 gave the best fit, which is an unusual setting.
- The method underperforms PSCCNet on Columbia, which the authors attribute to training-data distribution similarity.
Limitations
- Underperforms PSCCNet on Columbia dataset (1.4 AUC gap).
- Relies on a tension hyperparameter (tau=16) chosen empirically.
- Pre-training dataset is self-created and not released.
Experimental setup
Datasets: CASIA Columbia NIST16 COVER IMD20 RIFL21 synthetic tampering dataset (60K)
Benchmarks: Columbia Coverage CASIA NIST16 IMD20 RIFL21
Metrics: pixel-level AUC F1 score (EER threshold)
Baselines: SPAN TruFor PSCCNet ObjectFormer ManTra-Net RGB-N J-LSTM H-LSTM
Builds on: Mask R-CNN FCOS DeepLab ASPP Sobel layer
Closest papers by meaning
- Toward Real-world Text Image Forgery Localization: Structured and Interpretable Data Synthesis (NeurIPS 2025 · similarity 0.86)
- SUMI-IFL: An Information-Theoretic Framework for Image Forgery Localization with Sufficiency and Minimality Constraints (AAAI 2025 · similarity 0.85)
- Learning Discriminative Noise Guidance for Image Forgery Detection and Localization (AAAI 2024 · similarity 0.85)
- DiffForensics: Leveraging Diffusion Prior to Image Forgery Detection and Localization (CVPR 2024 · similarity 0.84)
- M²RL-Net: Multi-View and Multi-Level Relation Learning Network for Weakly-Supervised Image Forgery Detection (AAAI 2025 · similarity 0.84)
- FakeShield: Explainable Image Forgery Detection and Localization via Multi-modal Large Language Models (ICLR 2025 · similarity 0.84)
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.