Detection System Of Mask Wearing Based On FCOS Detection Algorithm

Novel coronavirus has been wreaking havoc all over the world since its outbreak,and has been also constantly mutating all the time.People must wear masks when they are in public places or closed spaces.Nowadays,the developing-rapidly target detection algorithm can detect whether a person wears a mask in real time,so that the situation whether people wear masks can be monitored.However,there are still some difficulties in the practical application of the algorithm.First of all,some people in densely populated scenes may not be detected due to the overcrowding.Second,when the target in in different positions and different light backgrounds,its size and definition will vary,which makes the algorithm difficult to detect.Then,the existing algorithm has a large model and a large consumption of calculation,which is not suitable to be applied on edge devices,mobile devices and other low computing power devices.How to reduce the model parameters while maintaining the original detection capability is also a difficulty in target detection.Therefore,the research on target detection task in practical application scenarios has practical and theoretical significance.In terms of the problem that people wearing masks in a closed environment such as food production workshop,this paper has made the improvement based on the FCOS detection algorithm.The contribution of this paper can be concluded as follows.First,this paper introduces Bi-FPN feature fusion algorithm to enhance the feature fusion of the model.GIOU is introduced as the regression loss function.In reality,due to the movement of the target and changes in light and angle,the images captured by the camera are blurry,overlapping,small and covered,which increase detection difficulty.The FCOS algorithm has a room for improvement in detecting the images in these scenarios,Bi-FN feature fusion can help the model improve the feature fusion capability of the low latitude feature layer,and GIOU loss function can strengthen the spatial sensitivity and overlapping sensitivity of the detection target.Second,in views of the sampling methods,this paper uses the central sampling.As data is inaccurate in the labeling process and the information near the label box is always negative samples,the inaccurate labeling information may be learned in the model training process.The central sampling allows the model to ignore the boundary information and pay more attention to the positive sample information during training,thus it effectively avoids the model from learning negative sample information.Third,this paper proposes to use Mobile Netv3 network as the feature extraction network,which significantly reduces the parameters of the model.At the same time,this paper uses the mask wearing dataset to conduct experiments.After using Mobile Netv3 features to extract the network,the accuracy of the model does not decline.In the meanwhile,it reduces the amount of calculation parameters by 43.7% and the size of the model by 69.1%.At last,the mask wearing detection algorithm that is proposed in the paper,combines with the Vue.js and the Spring Boot framework technology,design and realize a mask wearing monitoring system.It is applied to monitor the mask wearing in the production workshop of a food factory,which proves the effectiveness of the algorithm.