CatmullRom Splines-Based Regression for Image Forgery Localization

Li Zhang, Mingliang Xu, Dong Li, Jianming Du, Rujing Wang

Abstract

Image Forgery Localization (IFL) is critical for digital forensics, yet traditional segmentation-based methods often suffer from high false-positive rates and imprecise boundaries. This paper introduces CSR-Net (CatmullRom Splines-based Regression Network), a novel framework that reformulates IFL as a geometric regression task. By fitting adaptive CatmullRom splines to tampered regions and employing a dual-branch re-scoring mechanism, CSR-Net achieves sharper boundary localization and significantly reduces noise-induced errors. Experimental results across five benchmark datasets demonstrate state-of-the-art performance, particularly in handling complex, irregular forgery shapes.

The Challenge of Pixel-Wise Localization

Most existing IFL methods treat the problem as a semantic segmentation task, attempting to classify every pixel as either "authentic" or "forged." While intuitive, this approach faces two significant hurdles:

  1. Sensitivity to Thresholds: Segmentation models are often overly sensitive, flagging innocent regions as tampered due to lighting variations or sensor noise, leading to high False Positive (FP) rates.
  2. Boundary Inaccuracy: Forged regions typically have organic, irregular shapes. Segmentation masks often produce "blurry" or jagged edges that fail to align with the actual physical manipulation, while simple bounding boxes are too coarse for forensic precision.
Comparison of (a) segmentation-based, (b) standard regression-based, and (c) the proposed spline-based localization methods against Ground Truth (GT).

The CSR-Net Architecture

CSR-Net moves away from pixel-level classification in favor of continuous geometric regression. The architecture is built upon three core innovations designed to address the limitations of previous methods:

1. CatmullRom Spline Regression

Instead of a binary mask, the network predicts control points for CatmullRom splines. These interpolating splines are uniquely suited for IFL because they can adaptively fit complex, irregular contours much more effectively than standard polygons or quadrilaterals. A key finding was the importance of the "tension factor" ($\tau$); the researchers determined that $\tau = 16$ provides the optimal balance between fitting accuracy and curve smoothness, allowing the model to "hug" the contours of tampered objects.

2. Comprehensive Re-scoring Algorithm (CRA)

To mitigate the issue of false positives, CSR-Net utilizes a dual-branch prediction strategy. One branch provides a regional classification score, while the other provides an instance-level activation score. A candidate region is only confirmed as forged if it exhibits a strong response in both branches, effectively filtering out low-confidence "ghost" detections that plague traditional models.

3. Vertical Texture-interactive Perception (VTP)

Boundary refinement is handled by the VTP module, which employs learnable Sobel layers to extract horizontal and vertical texture features independently. By decoupling these orthogonal features, the model can resolve ambiguous edges that standard 2D convolutions might overlook, suppressing noise that would otherwise bleed across the forgery boundary.

The CSR-Net pipeline, highlighting the False Positive (FP) filtering via CRA and boundary refinement via VTP.

Experimental Results

CSR-Net was rigorously evaluated on five standard benchmarks: CASIA (v1 and v2), Columbia, NIST16, COVER, and IMD20. These datasets encompass a wide variety of manipulations, including splicing, copy-move, and object removal.

  • State-of-the-Art Accuracy: On the Coverage dataset, CSR-Net achieved a 97.9% AUC and a 78.0 F1-score, outperforming established models like PSCCNet and ObjectFormer.
  • Robustness: The model demonstrated high resilience to social media compression artifacts, which typically degrade the performance of pixel-sensitive segmentation models.
  • Efficiency: By regressing a limited set of control points rather than classifying millions of pixels, the method is computationally efficient and eliminates the need for heavy post-processing steps like Conditional Random Fields (CRF).

Conclusion

CSR-Net represents a paradigm shift in image forensics by treating forgery localization as a continuous geometric problem. The use of CatmullRom splines, combined with robust re-scoring and texture-aware refinement, allows for unprecedented precision in identifying tampered regions. Future work will explore the application of higher-order parametric curves to handle even more complex topologies, such as regions with internal holes, and the extension of this framework to real-time video forensics.

Summary updated

Abstract

IFL (Image Forgery Location) helps secure digital media forensics. However, many methods suffer from false detections (i.e., FPs) and inaccurate boundaries. In this paper, we proposed the CatmullRom Splines-based Regression Network (CSR-Net), which first rethinks the IFL task from the perspective of regression to deal with this problem. Specifically speaking, we propose an adaptive CutmullRom splines fitting scheme for coarse localization of the tampered regions. Then, for false positive cases, we first develop a novel re-scoring mechanism, which aims to filter out samples that cannot have responses on both the classification branch and the instance branch. Later on, to further restrict the boundaries, we design a learnable texture extraction module, which refines and enhances the contour representation by decoupling the horizontal and vertical forgery features to extract a more robust contour representation, thus suppressing FPs. Compared to segmentation-based methods, our method is simple but effective due to the unnecessity of post-processing. Extensive experiments show the superiority of CSR-Net to existing state-of-the-art methods, not only on standard natural image datasets but also on social media datasets.
BibTeX
@inproceedings{zhang2024catmullrom,
  author = {Li Zhang and Mingliang Xu and Dong Li and Jianming Du and Rujing Wang},
  title = {{CatmullRom Splines-Based Regression for Image Forgery Localization}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2024},
  doi = {10.1609/aaai.v38i7.28548},
}
Figure