EUVIMEDEUVIMED entdecken
Crossref

BC-IFL: A Multimodal Bias Correction Framework for Enhanced Image Forgery Localization

Haoyi Deng, Bin Li, Yuzhen Lin, Haodong Li, Changsheng Chen, Jiangqun Ni, Bo Cao

IEEE Transactions on Circuits and Systems for Video Technology · 2026-11 · 36 (11) · 17543-17556

DOI: 10.1109/tcsvt.2026.3711086 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Transactions on Circuits and Systems for Video Technology, 2026-11. Verfasst von Haoyi Deng und 6 weiteren Autor:innen. Die Arbeit nennt 47 Literaturangaben. Gefördert von: NSFC, Shenzhen Science and Technology Program.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

The rapid advancement of image editing and deep generative techniques has significantly increased the difficulty of verifying image authenticity. Although numerous Image Forgery Localization (IFL) methods have been proposed, most of them rely on fixed image-level thresholds to determine tampered regions, overlooking the inherent image-aware nature of pixel-wise logit predictions. Variations across images and within different regions of the same image often lead to diverse logit distributions, making unified thresholds suboptimal and limiting generalization. In this paper, we reformulate forgery localization refinement as a bias correction problem and propose a multimodal bias correction framework (BC-IFL) that introduces adaptive pixel-level thresholds for fine-grained decision making. Specifically, we fine-tune the visual encoder of CLIP via adapters to learn bias features supervised by the discrepancy between ground-truth masks and IFL model outputs, while simultaneously extracting auxiliary tampering features. These features are integrated by a Dynamic-weight Bias Correction Predictor (DBCP) to estimate pixel-level thresholds, where a dynamic weighting scheme adaptively balances different feature contributions during training. Furthermore, during the training phase, we introduce a Text Feature Enhancement (TFE) mechanism to augment the representational capacity of bias features and auxiliary tampering features via visual-text alignment, thereby enabling a more effective bias correction. Extensive experiments demonstrate that BC-IFL consistently improves the generalization and robustness of various state-of-the-art IFL models across multiple benchmarks.

Förderung

  • NSFC
  • Shenzhen Science and Technology Program

Zitieren

APA

Deng, H., Li, B., Lin, Y., Li, H., Chen, C., Ni, J., & Cao, B. (2026). BC-IFL: A Multimodal Bias Correction Framework for Enhanced Image Forgery Localization. IEEE Transactions on Circuits and Systems for Video Technology, 36(11), 17543-17556. https://doi.org/10.1109/tcsvt.2026.3711086

Vancouver

Deng H, Li B, Lin Y, Li H, Chen C, Ni J, et al. BC-IFL: A Multimodal Bias Correction Framework for Enhanced Image Forgery Localization. IEEE Transactions on Circuits and Systems for Video Technology. 2026;36(11):17543-17556. doi:10.1109/tcsvt.2026.3711086

BibTeX
@article{Deng2026,
  author = {Deng, Haoyi and Li, Bin and Lin, Yuzhen and Li, Haodong and Chen, Changsheng and Ni, Jiangqun and Cao, Bo},
  title = {BC-IFL: A Multimodal Bias Correction Framework for Enhanced Image Forgery Localization},
  journal = {IEEE Transactions on Circuits and Systems for Video Technology},
  year = {2026},
  volume = {36},
  number = {11},
  pages = {17543--17556},
  doi = {10.1109/tcsvt.2026.3711086}
}
Kostenlos starten →