In response to the issues of large model size, complex computations, and high resource demands on computational platforms in safety helmet detection models, a lightweight safety helmet detection algorithm called YOLOv8-MBS, based on an improvement of YOLOv8, is proposed. A new lightweight backbone module was first formed by combining MobileNetv3 with SPPF, reducing the algorithm′s parameter and computational load. Moreover, the algorithm′s feature extraction and representation capabilities are enhanced using a weighted bidirectional feature pyramid network, which also reduces the false detection rate. Finally, the SimAM module is incorporated to improve the network′s correlation between positional information and safety helmet features without increasing the computational burden. Experimental results show that compared to the original YOLOv8n network, the improved YOLOv8-MBS maintains high detection accuracy while reducing computation by 35.96%, the number of parameters by 25.63%, and model size by 23.22%, and increasing the frame rate by 12.52 fps. The lightweight nature of the model reduces deployment costs and provides theoretical support for embedded deployment and large-scale applications.