Mitigating Class Imbalance in PPE Compliance Object Detection via Data Augmentation

Personal Protective Equipment (PPE) crucially mitigates occupational injuries. While recent studies employ computer vision for PPE monitoring, the inherent foreground-foreground class imbalance problem remains frequently overlooked. To address this, we evaluate seven data augmentation (DA) policies across the imbalanced KSE-Head and SFCHD-4 datasets. Applying a combined DA policy improved YOLOv11's detection performance by 8.60% and 27.48% on KSE-Head and SFCHD-4, respectively. Furthermore, our findings demonstrate that techniques manipulating object scale, specifically Mosaic and Scale, contribute most significantly to enhancing PPE compliance detection.