Robust Image Manipulation Detection and Localization Using Cross-Domain Feature Learning
Mathematics · 14.11.2026 · 14 (22) · 4188
DOI: 10.3390/math14224188 ↗Kurz gefasst
Fachartikel in der Zeitschrift Mathematics, online erschienen am 14.11.2026. Verfasst von Ammar Odeh. Die Arbeit nennt 26 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Reliable image manipulation detection and localization remain challenging because deep learning models may learn dataset-specific characteristics rather than manipulation-related forensic evidence, reducing performance on unseen datasets and post-processed images. This study proposes Cross-Domain Feature Learning Network (CDFL-Net), a cross-domain feature learning framework for joint image-level manipulation detection and pixel-level localization. The method combines spatial image features with manipulation-sensitive residual representations. It applies class-conditioned cross-domain consistency, prototype stability, and inter-class separation objectives under label-preserving transformations, including Joint Photographic Experts Group (JPEG) compression, resizing, Gaussian blur, and additive noise. The framework was evaluated using within-dataset testing, zero-shot cross-dataset generalization, robustness analysis, ablation studies, and computational complexity measurements on public image-forensics benchmarks. On the held-out Chinese Academy of Sciences, Institute of Automation (CASIA v2) test set, CDFL-Net achieved a pixel-level F1-score of 0.908, Intersection over Union (IoU) of 0.831, image-level F1-score of 0.975, and image-level Area Under the Curve (AUC) of 0.995. Across CASIA v1, Columbia, Copy-Move Forgery Detection with Similar but Genuine Objects (COVERAGE), National Institute of Standards and Technology (NIST16), and Image Manipulation Dataset 2020 (IMD2020), it achieved an average cross-dataset pixel-level F1-score of 0.697. Robustness experiments also showed lower relative performance degradation under common post-processing operations. These results indicate that explicit cross-domain representation learning can improve the generalization and robustness of image manipulation detection and localization.
Zitieren
APA
Odeh, A. (2026). Robust Image Manipulation Detection and Localization Using Cross-Domain Feature Learning. Mathematics, 14(22), 4188. https://doi.org/10.3390/math14224188
Vancouver
Odeh A. Robust Image Manipulation Detection and Localization Using Cross-Domain Feature Learning. Mathematics. 2026;14(22):4188. doi:10.3390/math14224188
BibTeX
@article{Odeh2026,
author = {Odeh, Ammar},
title = {Robust Image Manipulation Detection and Localization Using Cross-Domain Feature Learning},
journal = {Mathematics},
year = {2026},
volume = {14},
number = {22},
pages = {4188},
doi = {10.3390/math14224188}
}
EUVIMED