Enhanced safety helmet detection with YOLOv8 and semisupervised learning via BiFPN and Wise-IoU v3

Safety helmets are crucial for ensuring worker safety in industrial environments. However, traditional object detection models often rely heavily on large, manually labeled datasets, which are both time-consuming and labor-intensive to produce. This paper presents an enhanced YOLOv8-based object detection model that leverages semi-supervised learning to reduce this dependency. Our approach incorporates Bidirectional Feature Pyramid Networks (BiFPN) for improved multiscale feature extraction and introduces the Wise-IoU v3 loss function for advanced bounding box regression. By implementing a self-training strategy, we generate pseudo-labels for unlabeled data, significantly mitigating the burden of manual annotation. Experimental results demonstrate a 1.3% improvement in mAP50 over the baseline YOLOv8 model, affirming the enhanced detection accuracy while also improving labeling speed to 10 milliseconds per image. This efficient and accurate solution offers promising prospects for safety helmet detection in real-world applications.