Rethinking Image Forgery Detection and Localization via Regression Perspective

Image Forgery Detection and Localization is rapidly advancing in the field of computer vision. Most methods locate forged regions in the form of segmentation and subsequently perform detection, facing challenges such as false detections (i.e., FPs) and inaccurate boundaries. In this work, we suggest rethinking the Image Forgery Detection and Localization (IFDL) task from a regression perspective and propose the CatmullRom Splines-based Regression Network (CSR-Net) to address these issues. Specifically, we first design an adaptive CutmullRom splines fitting scheme to predict coarse forged regions. Subsequently, we develop a novel rescoring mechanism that filters out samples with no response in both the classification and instance branches to reduce false positives. Besides, a learnable texture extraction module decouples horizontal and vertical forgery features, extracting more robust contour representations to further refine boundaries and suppress false detections. Compared to segmentation-based methods our method is simple but effective due to the unnecessity of post-processing. Extensive experiments conducted on several challenging benchmarks demonstrate that our method outperforms state-of-the-art methods qualitatively and quantitatively. Particularly, CSR-Net achieves optimal performance on three real-world datasets, indicating the applicability of our method to real scenarios such as social media and multi-tampered regions.