Research On Face Recongnition With Mask Based On Deep Learning

Since 2019,COVID-19 has swept the world,and it has become a routine phenomenon for people to wear masks to travel.Correctly wearing masks can effectively prevent the spread of viruses in the air and respiratory tract.Wearing masks in public places greatly reduces the risk of virus transmission.Masks become the first line of defense to prevent respiratory tract infectious diseases,which can effectively reduce the risk of novel coronavirus and H1N1 infection.By utilizing deep learning and artificial intelligence,machines can automatically detect whether a face is wearing a mask and identify the identity of the person wearing the mask,in order to meet the practical needs of face recognition in stations,airports,and other places without removing the mask under the normalization of the epidemic in the future.This algorithm system reduces staff contact and repetitive mask removal actions,avoids unnecessary trouble for epidemic management,and ensures the efficiency of epidemic prevention and control in important places.Wearing a mask facial recognition system can effectively prevent the spread of the epidemic and H1N1,ensure social safety and stability,improve work efficiency in places that require identity authentication,reduce the probability of cross infection,and save a lot of time and labor costs.The research and application of facial recognition systems with masks is of great significance in various fields such as computer vision,image processing,and deep learning.Its promotion and application can bring new opportunities and challenges to the development and innovation of related technologies,and promote technological progress.The research content is divided into the following parts:(1)The selection of facial detection and recognition algorithms.This article selects two algorithms for face detection and recognition with masks: the lightweight and fast detection network Mobile Net-SSD algorithm and the million level parameter Face Net face recognition algorithm.Compared to other detection algorithms,Mobile Net-SSD has the advantages of fast speed and fewer parameters in the field of object detection.In terms of large target face detection with masks,it has the characteristics of accuracy,recall,high m AP value,and fast speed.Facene is a conventional facial recognition algorithm with a deep backbone feature extraction network.Compared to other facial recognition algorithms,it has a simple acquisition path,strong universality,and a large number of parameters,resulting in high accuracy in facial recognition.(2)Improvements to the Mobile Net SSD algorithm for lightweight object detection networks.This paper builds a new feature extraction network Dyna Net by adding Mosaic data enhancement algorithm to enhance data and modifying the backbone feature extraction network Mobile Net network structure.The Smooth L1 Loss loss function is modified to Focal Loss,and based on this,the SSD algorithm(hereinafter referred to as Dyna Net SSD)is implemented to optimize the target detection efficiency.Through experiments,it has been proven that for small targets wearing masks,the detection accuracy has been improved by 6.9%,the recall rate has been improved by 4.3%,the AP value has been increased by 7%,and the m AP value has been increased by 6.82%;The accuracy rate for targets with a relatively large proportion has been increased by 1%,the recall rate has not changed,the AP value has been increased by 1.7%,and the m AP has been increased by 1.7%.Compared to the mainstream first stage object detection algorithms at present,the m AP value is lower than the mainstream detection algorithms by 4.3% and 2.6% for small and large targets,respectively,but the average speed is reduced by 50 ms.(3)Improve the facial recognition network Face Net.Due to the occlusion of facial feature information in facial images with masks,the available feature information is reduced,which leads to some problems in Face Net recognition,such as recognition errors and inability to recognize.This study adds online feature removal processing to the data,references the bottleneck residual module and ACNet convolution idea to modify the backbone feature network(Inception resnet-V1),and adds an improved CBAM block to form a new backbone feature extraction network,namely DAFace Net.Experiments have shown that the accuracy of facial recognition with masks is improved by 8.37% compared to Face Net,and the accuracy of facial recognition without masks is reduced by 0.1%.