Lightweight Safety Helmet Detection Algorithm Based on Improved YOLO11

Abstract

To enable the deployment of safety helmet detection on edge computing devices with limited resources, we propose a lightweight safety helmet detection algorithm, which is a safety helmet detection method based on YOLOv11n. Firstly, this method replaces the backbone network of YOLOv11n with the StarNet model to reduce the number of model parameters while extracting high-dimensional implicit features. Secondly, it adopts a multi-branch and multi-scale fusion module as the neck network (MBS-FPN). Through two bidirectional paths as the feature fusion network layer, and by repeating the same layer multiple times, higher-level feature fusion is achieved. In addition, shared convolution is introduced into the default detection head to reduce the number of model parameters and computational load, enhance the model's learning ability in multi-scale features, and improve the model's generalization ability in different backgrounds. Finally, the proposed lightweight safety helmet detection algorithm is verified on the public safety helmet dataset SHWD. The results show that compared with the original algorithm, the mAP50 and R are increased by 0.4% and 0.1% respectively, the number of parameters is reduced by 49.6%, and GFLOPs are decreased by 36.4%. In addition, by comparing the visual detection results of various algorithms, it is further proved that the lightweight safety helmet detection algorithm can provide a certain technical basis for the deployment of safety helmet detection.