Mask-wearing Detection System And Research In Complex Scenes

During the new type of coronavirus pneumonia(COVID-19)epidemic,mask-wearing detection is an important research hotspot in computer vision.In order to prevent the epidemic from recurring,it is very necessary for citizens to wear masks when going out during the epidemic.Therefore,it is also very urgent to apply a high-performance intelligent detection technology for wearing masks in public places,which can effectively reduce the supervision burden of staff,and reduce the probability of missed detection in crowded and crowded public places.Based on the deep learning method,this paper builds a mask-wearing detection system in complex scenes,and completes the detection tasks for both wearing and not wearing masks.This paper conducts research from the following three directions.(1)This paper proposes a improved object detection network based on the complexity of the mask-wearing detection scene.Specifically,a branch is established in the residual learning unit of the original object detection network RetinaNet,and an attention mechanism with feature recalibration characteristics is incorporated into the branch.The idea of this attention mechanism comes from the human visual system,the reflection in deep learning can be understood as acquiring the importance of each feature-map channel through feature learning,and then assigning weights to each channel based on this importance,so as to achieve the effect of enhancing useful information and suppressing useless or less useful information.Because the face occlusion is more serious in crowded scenes,the attention mechanism can be incorporated into the object detection network,and the local features of the object can be better learned by enhancing the weight of important information,so as to achieve better detection performance.(2)The application scenarios of mask-wearing detection technology are different,some are far and small objects,and some are near and large objects.Therefore,this paper further optimizes the network proposed in(1)to improve the detection performance of far and small objects in the scene.Specifically,by adding a shallow feature layer to the feature pyramid part of the network in(1),and integrating the shallow features into the deeper feature layer in turn.Compared with the deep features,although the shallow features have less convolution and weaker semantic information,but the resolution of the shallow features will be higher,and it also contains more detailed information related to location and texture.Therefore,adding shallow feature information is beneficial for the network to better learn the detailed information of far and small objects in feature learning.(3)In addition,this paper also uses random erasure and Mosaic data augmentation methods to expand the mask-wearing data-set,by increasing the number of occluded objects samples and enriching the background of the detected object to improve the generalization ability of the network.Finally,this paper builds a mask-wearing detection system based on the Pytorch experimental platform on the Ubuntu 16.04 operating system.Through the intelligent control system,we can complete daily mask-wearing detection tasks and intelligent release operations,and finally realize contactless intelligent supervision of mask-wearing during the epidemic.