Abstract: Face mask detection is a crucial component of intelligent surveillance systems and is of great significance in urban management and public health security. Aiming at the problems of missed detection and inaccurate detection that occur when dealing with images with issues such as lighting conditions and occlusion in mask face detection, the FSR-YOLOv8 face mask detection algorithm is proposed. This algorithm integrates an optimized version of the FasterNext backbone network to improve detection speed and reduce parameter quantity. Additionally, an improved SPCBAM (Spatial Pyramid Contextual Attention Module) attention mechanism is added between feature fusion points, allowing the model to more efficiently extract key features and minimize background interference. Furthermore, the construction of the RFB_HDC_BD (Receptive Field Block with Horizontal Dense Connections and Block Decimation) module significantly increases the utilization rate of feature semantic information. Under the same conditions, FSR-YOLOv8 outperforms YOLOv8 and other mainstream algorithms. Moreover, the model achieves a mean Average Precision (m_AP) value of 96.2% on the publicly available AIZOO dataset and an m_AP of 89.2% on a custom dataset, while the model parameter count is reduced by 16% compared to the YOLOv8 model. While the number of parameters was reduced, it had a relatively high accuracy.