Aiming at the problem of low accuracy of traditional face detection methods for large-volume mask-wearing people during the prevention and control of the new crown pneumonia epidemic, this paper proposes a real-time detection method for large-volume mask-wearing faces based on deep learning. The method uses the overall design of the backbone network, the FPN feature fusion network, the detection network and the parameter optimization method of the algorithm, and completes the model training on the mask-wearing face training set. In the detection process, the NMS algorithm is used to post-process the prediction results to realize multi-scale perception of the input face and improve the detection accuracy of the face wearing a mask. Experimentally verified, the detection accuracy of this method on the mask-wearing face test set is 0.919, and the average of Easy-0.841, Medium-0.802 and Hard-0.600 is obtained on the three subsets of the WIDER FACE test set. Detection accuracy (mAP). Compared with traditional face detection methods, it has universal advantages, and the video inference speed of the method in this paper reaches 55fps, which can meet the task requirements of real-time face detection with large traffic. In addition, the project team has successfully deployed this method to a fully automatic infrared thermal imaging temperature measurement warning system and put it into use in many places in Beijing, which is of great significance to preventing the spread of the epidemic.