Research On Mask Wearing Detection Algorithm In Dense Crowd Scenes

Since the outbreak of COVID-19,the country’s economic development and people’s daily life have been greatly affected.In order to prevent the spread of the epidemic,wearing masks in public places should be a necessary means of normalizing epidemic prevention and control.In crowded areas such as shopping malls and stations,manually checking the wearing of masks will consume a lot of human resources and be inefficient.Therefore,it is of great significance to use computer vision technology to efficiently and intelligently monitor the wearing of masks.At present,the mask wearing detection algorithm has the problems of weak anti-interference ability in dense crowd scenes,missed detection,false detection of occluded targets and small targets.In view of the above problems,this thesis studies the mask wearing detection algorithm based on YOLOv3.The specific work is as follows:(1)To solve the problem of insufficient utilization of associated feature information in mask wearing detection,an improved SE-YOLOv3 algorithm based on a multi-scale channel attention mechanism is proposed.The YOLOv3 feature extraction network is reconstructed through SENet,and the attention mechanism is used to mine the target context,thus making the network pay more attention to the associated target area.In addition,for the detection head of the YOLO series of algorithms,there is a space conflict between the classification task and the regression task,and the Decoupled Head is introduced on the basis of SE-YOLOv3.The classification and regression tasks are implemented separately through two branches and 8 convolution modules,and the two are integrated during prediction.The improved SE-YOLOv3-D algorithm optimizes the classification information and bounding box positioning information of the target in the feature layer,which further improves the detection accuracy.(2)Aiming at the problem that the lack of sample data in dense crowd scenes makes the model difficult to generalize,the Mosaic data augmentation method is used to expand the RMFD(Real-World Masked Face Dataset)dataset.Besides,the anchor box selection method is optimized through the K-means++ clustering algorithm to improve the model’s learning efficiency and recognition accuracy for masks and face targets.Then,the loss function is improved to speed up the convergence of the model and solve the problem of inaccurate positioning of the bounding box.The improved algorithm is named the YOLOMask algorithm.Through ablation experiments and robustness experiments,the improvement of the performance of the YOLO-Mask network brought by the aboveimproved strategies is more intuitively shown as a whole.Finally,the method in this thesis is compared with the current mainstream target detection algorithms.Experiments show that the average accuracy of the YOLO-Mask algorithm on the RMFD dataset reaches88.16%,which is 7.25% higher than the original YOLOv3.It can effectively reduce the missed detection rate and false detection of occluded targets,small targets,and targets in dense crowd scenes.Therefore,this algorithm can achieve high-precision real-time detection in the mask wearing detection task.