Towards Brightness-Robust Unified Anomaly Detection: Training with Multi-Brightness Data
Geonwoo Kim, Jimin Roh, Junho Lee, Suk‐Ju Kang · 2025
Real-world industrial environments often feature highly variable lighting conditions, causing performance drops in models trained solely under single-illumination data. To address this, we propose a multi-brightness data augmentation approach for One-for-All anomaly detection models, ensuring stable performance across diverse illumination environments. Specifically, we employ Test-Time Adaptation(TTA) from FiCo to generate images with varying brightness, reducing sensitivity to illumination changes. We evaluate performance using AUROC, Pixel-level AUC allowing fine-grained assessment beyond false-positive rates. Experimental results show that multi-brightness-trained models consistently maintain robust performance under novel lighting conditions, alleviating the need for frequent retraining and emphasizing the importance of real-world robustness in anomaly detection.