Improved Algorithm For Face Mask Detection Based On Yolov3

With the outbreak of the new crown epidemic in 2020,wearing a mask is one of the important measures to effectively suppress the rebound of the epidemic.It is of great practical significance to study the use of machine vision technology to detect whether a face is wearing a mask,but the performance of the existing mask wearing detection algorithm still needs to be improved.First of all,face detection images in complex scenes often contain occluded targets,small targets,targets with inconspicuous feature information,and targets close to the group.These targets are difficult to identify and detect;second,the existing mask wearing detection technology is classifying,Positioning accuracy and detection efficiency still cannot meet the requirements.In view of the above two problems,this paper divides the research content into two stages: single-target detection and multi-target detection.Two improved algorithms are proposed respectively,which both improve the detection performance of the model and have high application value.Specifically,in the single-target detection stage,a dataset of 3,000 faces wearing masks is produced,and the mask is labeled as "mask";in the multi-target detection stage,the dataset is expanded to 5000 and the labeling range is expanded to the entire face,and the mask is labeled as "mask".There are two categories: "mask" for faces wearing masks and "no＿mask" for faces without masks.Training and testing on the data set of the single target detection stage.Comparing the test results of the three algorithms of SSD,YOLOv3 and YOLOv4 show that YOLOv3 has the highest detection accuracy and speed,but it has the highest detection accuracy and speed,but it has the highest detection accuracy and speed.The detection effect of targets with inconspicuous information still needs to be improved.Therefore,a Single-YOLOv3 detection algorithm is proposed for the shortcomings of YOLOv3 in the single target detection stage.The algorithm replaces the Dark Net53 network with the designed DCN＿SERes PNet50 residual network,and DCN＿SERes PNet50 uses a small convolution.The function of the stacking and average pooling layer to achieve the purpose of improving the detection speed of the model and reducing the loss,combined with the DCN deformable convolution and the SENet channel attention mechanism,the model can better adapt to the face when wearing a mask due to occlusion The geometric deformation caused by other factors has enhanced the ability to express characteristic information.Experimental results show that the average accuracy of the algorithm proposed in the single target detection stage is as high as 95.36%,which is about 4.1% higher than YOLOv3,and the detection speed is 78.8FPS,which is 11.7FPS higher than YOLOv3.Train and test on the data set of the multi-target detection stage,and increase the comparison of SSD,YOLOv3,YOLOv4 and the Single-YOLOv3 algorithm proposed in the single-target detection stage.The test results show that SingleYOLOv3 has the best performance,but it still needs to be improved.Classification and positioning capabilities,therefore,a Multi-YOLOv3 detection algorithm is proposed for the shortcomings of Single-YOLOv3 in the multi-target detection stage.The algorithm first adds the Dropblock module to the feature fusion network to reduce the dependence between feature information and improve the model’s ability to classify different types of targets;secondly,the IoU-aware prediction branch is added to the detection head to strengthen the relationship between IoU and classification probabilities.The degree of association improves the positioning accuracy of the model.Experimental results show that compared with Single-YOLOv3,this method can effectively improve the detection performance of face wearing masks under multi-category conditions,and its detection accuracy in the "mask" and "no＿mask" categories is improved by about 1.2% and 3.2%,the average accuracy increased by 2.3%.