Aydın, Yıldız

This study presents a robust and descriptor-agnostic framework for copy–move forgery detection (CMFD) that combines handcrafted and deep features within a unified geometric refinement stage. The proposed approach integrates a lightweight alignment-based filtering (ABF) mechanism with multiple descriptors to reduce false positives left unresolved by RANSAC. Initially, ABF was evaluated with classical keypoint features Scale-Invariant Feature Descriptor (SIFT), Color-SIFT, and Color-RootSIFT on three benchmark datasets: CoMoFoD (200 images), MICC-F220 (220 images), and GRIP (80 forged + 80 authentic images). Compared with baseline RANSAC-only pipelines on the CoMoFoD dataset, ABF achieved consistent precision gains of 2%–5%, raising F1-scores from 0.811 to 0.834 (SIFT), 0.815 to 0.828 (Color-SIFT), and 0.854 to 0.861 (Color-RootSIFT). On the MICC-F220 dataset, the proposed Color-RootSIFT + ABF configuration reached an F1-score of 0.888, surpassing several recent deep-learning methods (0.84–0.85). To further validate its generality, ABF was combined with a Multiscale Color-R2D2 deep descriptor that extracts LAB-channel features across three scales. This configuration achieved a perfect F1-score of 0.9937 on the GRIP dataset and an F1-score of 0.9149 on the Image Manipulation Dataset (IMD, 48 forged and 48 authentic images). The obtained results are competitive with those of recent deep-learning–based CMFD methods. Moreover, the proposed framework requires no retraining and substantially lower computational resources. These results demonstrate that ABF provides a simple yet powerful geometric verification layer, offering a transparent, computationally efficient, and interpretable alternative to complex deep-learning–based CMFD frameworks.