Research On Mask Wearing Detection Method In Complex Scenes Based On Deep Learning

Under the invasion of Derta and Omickrong variant strains,the epidemic situation is complex and changeable,and there is a possibility of outbreak at any time.With the continuous breakthrough of computer performance,target detection technology has been developed rapidly.Under the normalized epidemic situation,mask wearing detection technology has become a new application direction in the field of target detection.However,under the influence of different scale changes,overlapping occlusion and perspective in complex scenes,mask wearing detection and recognition is difficult,with high missed detection rate and slow detection efficiency.In view of the above problems,this paper optimizes the YOLOv3 target detection algorithm based on deep learning,and studies the problem of mask wear detection in complex scenes.The main research contents are as follows:Firstly,aiming at the problem of poor detection effect caused by density,occlusion and small size targets in mask wearing detection task in complex scenes,a high-precision improved YOLOv3-M network is proposed.Firstly,based on the original network structure,the shallow features are fused to construct the shallow detection layer to form a four-scale detection structure to improve the accuracy of small target detection.Then,a multi-scale fusion network combined with path enhancement strategy is introduced to further utilize the feature information to achieve feature enhancement;then CIo U loss is selected to optimize the border regression to improve the positioning accuracy;finally,K-Means algorithm is used to re-cluster the data set to improve the adaptability of the prior frame.The experimental results show that the average accuracy of the proposed method reaches 93.66%,which is 5.61% higher than that of the original YOLOv3 algorithm,effectively reducing the missed detection rate and improving the detection effect of the algorithm on the target in complex scenes.Secondly,aiming at the problems of complex structure and large parameters of YOLOv3 algorithm model,a lightweight improved YOLOv3-MS network is proposed.Firstly,the cross-stage local network structure and Ghost Module module are introduced into the backbone network to greatly reduce the network convolution parameters and reduce the computational complexity.Then,the deep separable convolution is used to replace part of the standard convolution in the network to further reduce the number of model parameters;finally,the improved pyramid pooling structure is added after the backbone network to enrich the extracted feature information.The experimental results show that the proposed method can effectively reduce the complexity of the model,improve the detection efficiency and ensure high detection accuracy.The average accuracy rate reaches 92.53%,which is 4.48% higher than the original YOLOv3 algorithm.At the same time,the model size is reduced by 76%,the detection rate is increased by 11.7 frames/s,and the FPS reaches 47.4.It realizes the detection and recognition of mask wear accurately and efficiently,and has strong robustness and generalization ability in complex scenes.Finally,in order to facilitate the operation of supervisors,increase the humancomputer interaction experience,and visualize the recognition results intuitively,this paper designs and develops a software system based on Python-Tkinter for the detection algorithm to realize the detection of local images and video files and the real-time detection of cameras.It has good practicability and application prospect for the current severe epidemic prevention situation.Figure 61,table 11,reference 63.