Wearing masks is an effective way for preventing COVID-19 and cooperating with the national epidemic prevention and control. An improved YOLOv7 algorithm is proposed to solve the problems such as whether masks are correctly worn, different shooting angles and being blocked. Based on YOLOv7, the convolutional attention mechanism is introduced into the Head region of the network to make the feature network more targeted in the processing of the mask region, thus enhancing the learning ability of the feature network to the mask region. The structure of Backbone area is optimized, the ConvNeXt network structure is improved, and partial convolution is introduced into the network instead, which improves the detection accuracy and robustness of the model and enhances the accuracy of prediction without introducing a large number of additional calculations. The space pyramid pool of the Head layer is improved to improve the training speed and accelerate the model convergence. Experiments show that in the case of complexity and occlusion, the loss function of the improved YOLOv7 decreases significantly, and the mAP on the test set is 93.8%, which is 3.6% higher than that of the original YOLOv7 algorithm.The accuracy of each category is improved, and the accuracy of no mask, correct mask and incorrect mask are increased by 6.8%, 2.1% and 1.7%, respectively. The cases of error detection are significantly reduced, and the generalization ability is significantly improved.