Research On Mask Wearing Detection Algorithm Based On Improved YOLOv3

The new type of coronavirus can be spread by carriers such as droplets and aerosols in the air and is extremely infections.Consciously wearing masks in public places is the most effective measure to prevent a rebound of the epidemic.In addition,some special workplaces(such as polishing workshops and dust-free workshops)require workers to wear masks at all times for work.For the situation of wearing a mask,manual patrol investigation is time-consuming and laborious.Therefore,it is a trend to use a camera to perform intelligent monitoring on the PC side.At present,most mask wearing detection algorithms only focus on detection accuracy,and the detection efficiency on the low-performance PC side is not ideal.For this reason,this topic starts from judging whether a face is wearing a mask,based on the idea of the YOLOv3 algorithm,and intends to explore a lightweight mask wearing detection algorithm with high detection accuracy and fast speed.First of all,to meet the requirements of light weight and high speed for mask wearing detection,feature extraction network MobileNetV2 based on depthwise separable convolution and inverted residual structure was introduced in the main network of YOLOv3.The standard convolution is divided into depthwise convolution and pointwise convolution,which greatly reduces the number of network parameters and computational complexity,thus improving the detection speed of the algorithm.Through the inverted residual structure,the dimension of the channel first rises and then falls,enriching the feature information extracted by the depthwise separable convolution.Secondly,the introduction of the spatial pyramid pooling structure after the feature extraction network realizes the fusion of global features and local features,enriches the expression ability of the final feature map,and improves the detection accuracy of the model without increasing the computational complexity of the network.Finally,the training process of the network is optimized by improving the regression loss of the bounding box and data augmentation.YOLOv3 uses the L norm to calculate the regression loss of the bounding box during the training process,and uses the intersection-over-union to determine whether the target is detected during the evaluation process.There is an inequality problem between the two.Therefore,this paper optimizes the regression loss of the bounding box to the loss function related to the intersection-over-union.In addition,this paper uses the Mosaic data augmentation method to further improve the generalization ability and anti-interference ability of the network.The method in this paper is evaluated by cross-validation on a self-built dataset.The experimental results show that the method in this paper greatly reduces the complexity of the network while ensuring detection accuracy.The detection speed can reach 125 frames per second on the experimental platform,which has an absolute speed advantage compared with several excellent detection algorithms.At the same time,the method presented in this paper still has good robustness under the conditions of brightness,scale change and occlusion,and has certain adaptability to real-time detection in complex environments.