Research on real-time helmet detection and deployment based on an improved YOLOv7 network with channel pruning

Wearing helmets correctly is crucial for the safety of workers in industrial and construction settings. This paper introduces an improved YOLOv7 algorithm that is designed to achieve efficient real-time detection of the usage and deployment of safety helmets. The improved YOLOv7 algorithm utilizes ReXNet as its backbone network and enhances the feature extraction ability of the constructed model by modifying the feature output layer of the network. To address the challenge of small target detection, the P2 detection layer is added to the backbone network to mitigate the loss of small target features. Furthermore, an asymptotic feature pyramid network (AFPN) is introduced in the neck part to facilitate direct interaction among the nonadjacent layers and feature fusion. Additionally, the channel pruning algorithm is applied to simplify the improved YOLOv7 detection model, which significantly reduces the number of model parameters, the model size, and the number of floating-point operations (FLOPs) by 74.5%, 73.4%, and 52.0%, respectively. The size of the pruned model is only 19.9 MB. Compared with 8 mainstream algorithms, this algorithm has better performance in terms of both accuracy and efficiency. Finally, by deploying the trained model on edge development equipment, validating the effectiveness of this helmet detection algorithm at industrial sites. In summary, the proposed lightweight helmet detection algorithm based on an improved YOLOv7 network satisfies the real-time requirements imposed in the field, providing technical support for safety inspection tasks in complex industrial environments.