Deep learning is usually used for face mask detection. The algorithm based on one-stage mAP is usually around 94% or 96% in current approaches. Although mAP is around 98% in two-stage algorithm-based approaches, the weight is large and the FPS is poor which is not suitable for face mask detection. To address these issues, this paper we propose a lightweight face mask detection method based on improved YOLOv5 and a new calibration method. We choose YOLOv5 to improve convinced that our model also has advantages over the latest YOLOv8. Our model first replaces the C3 module with a new partial convolution to achieve a lighter network and decrease the redundant parameters; Secondly, a bi-directional feature pyramid network(BiFPN) neck network is designed to enhance the feature extraction capability of the algorithm and enrich the semantic information; Thirdly, we combine the SE attention with C3 module to increase adaptability and robustness; Finally, the EIoU loss function and SGDR optimization algorithm are employed to accelerate convergence speed and improve generalization ability. Experimental results on the face mask datasets demonstrate that our model compared to the original YOLOv8s, achieving a 48% reduction in computation, 46% reduction in parameters, 44% reduction in weight, and 98.7% mAP exceeds most other models. And our calibration method has also been demonstrated to be effective in enhancing the mAP, these results validate the effectiveness of the proposed method.