EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

See What’s Wrong: Visual Counterfactual Explanations for Anomaly Detection via Energy-Guided Restoration

Eirini Cholopoulou, Dimitris K. Iakovidis

IEEE Open Journal of Signal Processing · 2026 · 7 · 822-832

DOI: 10.1109/ojsp.2026.3712191 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Open Journal of Signal Processing, 2026. Verfasst von Eirini Cholopoulou und 1 weiteren Autor:innen. Die Arbeit nennt 43 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: European Union (EU)’s Horizon Europe research and innovation, Open Access funding supported by HEAL-Link.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Visual Anomaly Detection (VAD) aims to identify patterns in images that deviate from normality in real-world applications, such as industrial quality inspection and medical diagnosis. While recent generative and reconstruction-based methods have been proposed in the literature, they remain prone to hallucinations, altering or over-correcting normal regions without explicit guidance. Furthermore, these approaches operate as black boxes, which limits user trust. To address these limitations, this paper proposes a novel methodology that regularizes the generative restoration process using a learned signal prior. Unlike traditional methods that blindly restore the entire input and provide only anomaly scores, the proposed approach combines Energy-Based Modeling with diffusion-guided inpainting to generate counterfactual explanations for VAD. More specifically it introduces a lightweight EBM that utilizes a contrastive energy loss to associate high-energy values with potential abnormalities, producing a spatial prior that subsequently constrains the restoration process only to potential anomalies, preserving normal image context and minimizing hallucinations. The restored images then serve as intuitive counterfactual explanations that illustrate how a sample would appear if it were normal, enabling a direct pixel-level comparison between the input and the energy-guided output. Experiments on the MVTec AD and VisA datasets demonstrate that the proposed method achieves an average pixel-level localization AUROC score of 97.5% and 97.3%, respectively, outperforming state-of-the-art reconstruction methods while providing visual counterfactual explanations that preserve the original context.

Förderung

  • European Union (EU)’s Horizon Europe research and innovation
  • Open Access funding supported by HEAL-Link

Zitieren

APA

Cholopoulou, E., & Iakovidis, D. K. (2026). See What’s Wrong: Visual Counterfactual Explanations for Anomaly Detection via Energy-Guided Restoration. IEEE Open Journal of Signal Processing, 7, 822-832. https://doi.org/10.1109/ojsp.2026.3712191

Vancouver

Cholopoulou E, Iakovidis DK. See What’s Wrong: Visual Counterfactual Explanations for Anomaly Detection via Energy-Guided Restoration. IEEE Open Journal of Signal Processing. 2026;7:822-832. doi:10.1109/ojsp.2026.3712191

BibTeX
@article{Cholopoulou2026,
  author = {Cholopoulou, Eirini and Iakovidis, Dimitris K.},
  title = {See What’s Wrong: Visual Counterfactual Explanations for Anomaly Detection via Energy-Guided Restoration},
  journal = {IEEE Open Journal of Signal Processing},
  year = {2026},
  volume = {7},
  pages = {822--832},
  doi = {10.1109/ojsp.2026.3712191}
}
Kostenlos starten →