WeakLoc: Low-Rank Attention and Adaptive Pseudo-Label Learning for Weakly Supervised Image Forgery Localization
IEEE Transactions on Circuits and Systems for Video Technology · 2026-11 · 36 (11) · 16978-16989
DOI: 10.1109/tcsvt.2026.3717773 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Transactions on Circuits and Systems for Video Technology, 2026-11. Verfasst von Ningning Bai und 4 weiteren Autor:innen. Die Arbeit wurde bisher 1-mal zitiert (Crossref) und nennt 58 Literaturangaben. Gefördert von: National Natural Science Foundation of China.
Abstract (Originalsprache)
With the rapid advancement of generative artificial intelligence, image forgery has become increasingly realistic and accessible, posing significant challenges to digital media forensics. Existing image forgery localization methods largely rely on pixel-level annotations, whose expensive acquisition limits the scalability of supervised learning and prevents the effective exploitation of large-scale weakly labeled data. To address this limitation, we propose WeakLoc, a weakly supervised image forgery localization framework that combines low-rank attention with adaptive pseudo-label learning. Specifically, WeakLoc first reconstructs the input image using an implicit neural representation and derives a high-frequency discrepancy map by comparing the reconstructed image with the original one, thereby exposing subtle manipulation artifacts that are difficult to capture from semantic features alone. Based on these forensic priors, an Omnidirectional Suspicious Clue Extractor (OSCE) aggregates manipulation evidence across multiple semantic scales to produce discriminative class activation maps for coarse localization. To improve localization accuracy under weak supervision, we introduce a Forgery–Background Decoupling Module (FBDM), where low-rank attention explicitly exploits the low-rank property of authentic background features to suppress redundant background responses and enhance the separability of forged regions. Furthermore, within a teacher–student learning framework, the teacher progressively refines localization predictions by dynamically integrating class activation maps with its own predictions to generate adaptive pseudo labels, which provide increasingly reliable supervision for training the student network. Extensive experiments on multiple public image forgery benchmarks demonstrate that WeakLoc consistently outperforms existing weakly supervised methods and achieves performance competitive with fully supervised approaches.
Förderung
- National Natural Science Foundation of China
Zitieren
APA
Bai, N., Wang, X., Hou, J., Wang, Q., & Liu, Z. (2026). WeakLoc: Low-Rank Attention and Adaptive Pseudo-Label Learning for Weakly Supervised Image Forgery Localization. IEEE Transactions on Circuits and Systems for Video Technology, 36(11), 16978-16989. https://doi.org/10.1109/tcsvt.2026.3717773
Vancouver
Bai N, Wang X, Hou J, Wang Q, Liu Z. WeakLoc: Low-Rank Attention and Adaptive Pseudo-Label Learning for Weakly Supervised Image Forgery Localization. IEEE Transactions on Circuits and Systems for Video Technology. 2026;36(11):16978-16989. doi:10.1109/tcsvt.2026.3717773
BibTeX
@article{Bai2026,
author = {Bai, Ningning and Wang, Xiaofeng and Hou, Jianpeng and Wang, Qin and Liu, Zinian},
title = {WeakLoc: Low-Rank Attention and Adaptive Pseudo-Label Learning for Weakly Supervised Image Forgery Localization},
journal = {IEEE Transactions on Circuits and Systems for Video Technology},
year = {2026},
volume = {36},
number = {11},
pages = {16978--16989},
doi = {10.1109/tcsvt.2026.3717773}
}
EUVIMED