Intelligent Detection Algorithm On Mask Wearing

Since January 2020,Because of the rapid spread of COVID-19,Both health of people all over the world and overall economic level have been greatly affected.Wearing masks can effectively prevent the spread of Covid-19.So,in stations,shopping malls,public transport vehicles and other public places,manual detection points are usually set at the entrance to remind and check the wearing of masks,measure the body temperature and check the health QR code.Because of the large population base in China,the work pressure of detection personnel is greatly increased.Especially for the long-term monitoring of mask wearing in a closed environment,the manual detection method is not competent.Therefore,automatic detection of mask wearing has become a new epidemic prevention demand.In view of the challenges faced by mask wearing detection in public places,such as real-time,multi-target,multi-posture and facial occlusion,this paper studies the mask wearing detection method with high detection rate and real-time.The main work is as follows:(1)In view of the lack of mask wearing detection dataset,this paper establishes a mask wearing detection dataset containing 5127 pictures,which provides a data basis for model training and model effect test in the follow-up research work.The dataset covers a variety of public and complex scenes at home and abroad such as airports,stations and subways,including a variety of mask styles and occlusion,which can ensure the diversity of data,and Label Img is used for annotation according to the requirements of model detection.(2)Designed mask wearing detection method based on YOLOv4.On the one hand,by introducing the ideas of void convolution and asymmetric convolution,combined with RFB method,the feature enhancement module RFB-s is designed,and the spatial pyramid structure in YOLOv4 is replaced,which expands the receptive field of the model and reduces the amount of network parameters.On the other hand,the attention module is added to improve the information processing ability of the model.Through experiments on self built Mask Dataset and Mask Open Source Dataset,the m AP value and FPS under different network structures and different algorithms are compared to verify the improvement of the algorithm in this chapter in mask wearing detection performance.(3)According to the real-time requirement of surveillance video detection,an improved method of mask wearing real-time is further designed.By introducing the lightweight network Mobile Net-v3 as the backbone and adding the attention module,the model information processing ability is improved.At the same time,void convolution and CSPNet are introduced to modify the network prediction part,enhance the feature extraction ability of the model and reduce the model parameters.The improved model maintains the lightweight characteristics,improves the accuracy to a certain extent,and greatly reduces the model training and detection time.It is more suitable for the detection of surveillance video and helps to detect the wearing of masks in a specific space in real time.The experimental results show that the average detection accuracy of rfbnet-s on two different datasets is 99.38% and 99.13% respectively;The average detection accuracy of the mask wearing real-time improvement method Yolo mask on two different datasets is 95.6%and 92.3% respectively,while the FPS value is 67.6 and 65.8 respectively,which is 30.4 and29.3 higher than the first detection method.The two methods basically solve the problems of low detection rate and low real-time performance faced by the mask wearing detection method at this stage.