Enhancing Low-Light PPE Violation Detection in Industrial Environments Using Multi-Contrast Image Processing and YOLOv9

Abstract

Ensuring compliance with Personal Protective Equipment (PPE) is critical for workplace safety, yet low-light conditions reduce the accuracy of vision-based monitoring systems. This study introduces a novel violation-aware PPE dataset consisting of 11,407 annotated images across eight categories, explicitly modeling both compliance and non-compliance classes (e.g., helmet/no-helmet, vest/no-vest, gloves/no-gloves, shoes/no-shoes). The dataset differs from prior works by focusing on low-light industrial environments where detection is most challenging. To address these conditions, three multi-contrast enhancement techniques—Contrast Limited Adaptive Histogram Equalization (CLAHE), Auto-CLAHE, and EnlightenGAN—were integrated with an optimized YOLOv9t model. The six You Only Look Once (YOLO) variants (YOLOv5s–YOLOv12n) were benchmarked, with YOLOv9t plus augmentation achieving the best performance, with mAP@50 of 0.915, mAP@50–95 of 0.659, precision of 0.906, recall of 0.844, and inference time of 4.0 ms. The enhancement experiments demonstrated that CLAHE provided the highest detection coverage (79.43%), Auto-CLAHE yielded the greatest detection density (2.29/frame), and EnlightenGAN offered limited benefits due to domain shift. The findings confirm that histogram-based methods consistently improve PPE violation detection under low-light conditions, while Generative Adversarial Network (GAN)-based approaches require domain-specific adaptation. Overall, this study contributes a new dataset, systematic YOLO benchmarking, and the first task-driven comparison of classical, adaptive, and GAN-based enhancement methods for reliable, real-time workplace safety in challenging lighting environments.