Image-based anomaly detection in low-light industrial environments with feature enhancement
Industrial anomaly detection and localization are essential for maintaining product quality and safety in manufacturing. However, these tasks become significantly more challenging in low-light environments, where poor illumination introduces noise and reduces visibility, leading to degraded performa...
Saved in:
| Main Authors: | , , , , , , , , , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2025-03-01
|
| Series: | Results in Engineering |
| Subjects: | |
| Online Access: | http://www.sciencedirect.com/science/article/pii/S2590123025003901 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Industrial anomaly detection and localization are essential for maintaining product quality and safety in manufacturing. However, these tasks become significantly more challenging in low-light environments, where poor illumination introduces noise and reduces visibility, leading to degraded performance of traditional methods. Existing approaches often rely on computationally expensive low-light image enhancement techniques or are limited by their sensitivity to noise and inability to adapt to varying illumination conditions, resulting in suboptimal anomaly detection performance. In this paper, we present DarkAD, a novel end-to-end approach specifically designed to address the difficulties of anomaly detection under low-light conditions. Our framework features a Dark-Aware Feature Adapter (DAFA), which enhances feature extraction by integrating two key modules: Frequency-based Feature Enhancement (FFE) and Illumination-aware Feature Enhancement (IFE). The FFE module suppresses high-frequency noise and amplifies structural details, while the IFE module adaptively boosts features from both well-lit and dimly illuminated regions, allowing the model to focus on critical areas without relying on computationally expensive image enhancement techniques. Extensive experiments on multiple industrial object categories demonstrate that DarkAD significantly outperforms state-of-the-art methods. Specifically, it achieves a mean image-level anomaly detection accuracy (I-AUROC) of 0.899 and a localization performance (AUPRO) of 0.862, surpassing the next-best method by 16.2% and 16.5%, respectively. Additionally, the framework maintains a real-time inference speed of 60 frames per second (FPS), making it well-suited for deployment in industrial settings. The ablation study further highlights the synergistic contributions of FFE and IFE, with their combined effect driving the model's superior performance across diverse and complex scenarios. |
|---|---|
| ISSN: | 2590-1230 |