💻 Developed through an ICT Vision AI Project

🎉 Our model was accepted at CVPR 2025 Workshop

📌 Official Implementation code: SK-RD4AD

This research focuses on improving the robustness and localization ability of one-class anomaly detection models by introducing non-corresponding skip connections into a reverse distillation framework.

To address the limitations of deep-layer feature loss in conventional reverse distillation frameworks, we propose a novel skip-connected design that enhances multi-scale feature preservation and localization accuracy.

Evaluated on MVTec-AD, Valeo Anomaly Dataset (VAD), and VisA.

Our experiments demonstrate that SK-RD4AD significantly improves anomaly localization, particularly in datasets with complex textures (e.g., VisA) or subtle defects (e.g., VAD).
The introduction of non-corresponding skip connections enables better flow of spatial detail into deeper layers, allowing the model to detect fine-grained anomalies more precisely.

• On VisA, the Pixel AUPRO improved by +21.2, showing enhanced ability to localize structural anomalies in PCB data.
• On VAD, the model showed better generalization across real-world driving scenarios, achieving +2.5 AUROC gain.
• Even on MVTec, where performance is already strong, the model shows consistent gains, indicating stability.

These results validate the hypothesis that feature loss in deep layers is a core limitation in prior work, and that skip-connected reverse distillation can effectively address this issue.

Anomaly maps clearly highlight defective regions with high accuracy.
Red/yellow hues indicate high anomaly confidence, overlayed on the original images for interpretability.

• Non-Corresponding Skip Connections: Enhances deep layer representations with fine-grained details.
• Robust Generalization: Performs reliably across various industrial and structural defect types.
• Built on Reverse Distillation: Based on the RD4AD backbone, improved for feature preservation.