Research On Mask Wearing Detection Model Based On Improved YOLOv3

The COVID-19 pandemic has posed a great threat to human life and health,with the pandemic spreading all over the world,wearing masks has been considered as an effective measure to prevent COVID-19 from spreading.At present,the detection of mask wearing is mainly based on manual monitoring.On the one hand,it costs manpower,on the other hand,it will cause people queuing and gathering because of the low efficiency of manual monitoring,which increases the risk of virus transmission in turn.There are very few mask detection methods on the basis of computer vision,inevitably,there are some problems such as weak anti-interference ability,small target missing detection,false detection and slow detection speed in complex background.In order to solve the above problems,in this dissertation,deep learning method was adopted to detect the mask wearing on the basis of computer vision.the specific research contents are as follows:Firstly,the representative target detection and recognition model Faster R-CNN and YOLOv3 were selected as the initial models of mask wearing detection,subsequently,the training was carried out based on the self-made mask data set,and the detection performance of the two models was compared simultaneously.Finally,the YOLOv3 model demonstrated more satisfying detection accuracy and speed and was selected as the benchmark model for subsequent optimization and improvement.Secondly,in order to overcome the shortcomings of the existing models,such as imprecise bounding-box regression and tedious forward chaining calculation,in this dissertation,the optimization strategy of anchors and basic components BDL was employed.The results showed that the optimization strategy of anchors and basic components not only improved the fitting degree between the detected frame and the target,but also improved the forward chaining speed.In detail,compared with the benchmark model,the detection accuracy was 85.5% which increased by 2.1%,while the detection speed was 0.02 s/sheet which increased by 6.4%.Thirdly,in order to solve the problems of low detection accuracy,small target missing detection and false detection in the existing model,in this design,an improved network integration of YOLOv3 and Dense Net was established,specifically,the Dense Net module was adopted to replace the Res Net module for extracting small and medium scale features,subsequently,the target in this size was applied to train the model.The results revealed that the improved strategy on the basis of the Dense Net improved the sensitivity of the model to small and medium targets,with improving the missed detection rate and false detection rate,the detection accuracy of the model was 88.8% which was improved by 3.3%.Fourth,in order to further make up for the defects of low detection accuracy and weak target searching ability,in this design,the spatial transformer network module was added to the backbone network,by which the model was able to accurately focus on the more pivotal location feature information.The results illustrated that the improved strategy based on attention mechanism improved the search ability of the model for small targets and targets that are difficult to detect,with further improving the miss detection rate,the detection accuracy was 90.7% which increased by 1.9%.Finally,in order to solve the problem that the existing models are vulnerable to background interference in complex scenes,in this design,an adaptive multi-scale feature fusion network structure was established,each feature map employed for prediction contained three scales of feature information,and the information expression of target feature was enhanced by redistributing the weight of feature channel.The results demonstrated that the improved strategy on the basis of adaptive feature fusion was capable of empowering the model to detect the target more accurately in various complex environments.Notably,while ensuring the high detection speed,the detection accuracy reached 93.4% which increased by 2.7%.