Research on detection method of mask wearing based on improved YOLOv5

The network model of existing mask wearing detection methods have large number of parameters and calculation. To improve these shortcomings, this paper proposed a lightweight mask wearing detection method based on improved YOLOv5. This method replaced the original convolution module by Ghost module reducing the model parameters and computation. This method used lightweight attention mechanism and Transformer self-attention module to make the model more focused on important information to improve accuracy. This method used Alpha-CIoU loss function to enhance the accuracy of bounding box regression. At last, this method used knowledge distillation to boost the detection performance of small model. The results of the experiment show that this method reduced the number of parameters, computation amount and model size to 57.3%, 51.83% and 58.54%, and improved the mAP by 0.7%.