EL-FDL: Improving Image Forgery Detection and Localization via Ensemble Learning

The widespread dissemination of diverse forgery images has profoundly impacted social life. Thus, image forgery detection techniques are becoming increasingly urgent. Existing models are usually trained to detect certain types of forgery images, leading to an insufficient generalization in detecting various forgery images (e.g., copy-move, splicing, inpainting). In this paper, we conducted extensive testing on SOTA models and revealed the limitations of individual models, including 1) insufficient generalization capability and 2) high false positive rates for pristine images. To address the above issues, we propose EL-FDL, a method based on stacking ensemble learning, which enhances the detection and localization abilities by integrating the output of heterogeneous SOTA models. Extensive experimental results demonstrate that our proposed EL-FDL significantly improved: +16.4% in detection, +11.1% in localization, and overall false positive rate decreased by at least -21.0% across the test dataset.