Mask Detection Algorithm Based On Improved YOLOv4-tiny

Since the massive spread of Corona Virus Disease 2019(COVID-19)in December2019,the virus has spreaded more than 200 countries and regions around the world,posing a major threat to life and health of people all over the world.COVID-19 can be acquired by droplets transmission and contact transmission mainly.Generally,monitoring temperatures,wearing masks and keep a distance in public places could block the transmission of the virus and curb the spread of the epidemic.Public transport facilities such as airports,stations and so on,due to the special nature of functions and high density of crowd,are important places for COVID-19 controlling.Crowd regulation in such public places is one of the important tasks during the period of COVID-19 controlling.In order to efficiently monitor the coverage of mask wearing so that the controlling personnel could take measures in a timely manner,this thesis proposes a mask detection algorithm based on improved YOLOv4-Tiny.The main work of this thesis is as follows:1.A mask detection dataset containing 13,324 images and detailed annotations was generated for model training and testing of mask detection task.The images of the dataset are distributed in a variety of real social scenes and annotated in detail in conjunction with the relative requirements of the mask detection task.2.A spatial pyramid pooling module is developed after the backbone of YOLOv4-tiny.This module pools and fuses the input feature map at multi-scale to enhance the receptive field of the network and increase the amount of information of feature map.3.Use the path aggregation network instead of the feature pyramid network of the YOLOv4-tiny as the feature enhancement module.The input feature maps of different scales are divided into bottom-up and top-down paths for feature fusion and repeated enhancement to make the best of the detailed information extracted from the backbone network.4.Combining the CIo U loss function and label smoothing strategy,the location loss function and predicted category loss function of bounding box in mask detection are optimized.And using the Mosaic data-augmentation method and the cosine annealing learning rate schedule to improve the training efficiency and detection accuracy of the model.5.In the ablation experiment,several control trials were setted by controlling the relevant variates,which verified the effectiveness of improvement on network structure and training strategies.The experimental results show that the detection accuracy of the proposed algorithm on the mask target and face target is 94.7% and 85.7%,respectively,which is 4.3% and 7.1% higher than that of the YOLOv4-tiny.In the meanwhile,the real-time detection speed achieves 76.8 FPS(tested on GTX 1050 Ti GPU).The proposed algorithm satisfies the accuracy and real-time speed of mask detection tasks in various scenes.