Research On Face Recognition Algorithm Under Local Occlusion

With the progress of society,the development of science and technology,and the accumulation of data,face recognition technology has become more mature,and has been widely used in daily life scenarios such as station security inspection,campus access control,scenic spot ticket checking,face-scanning payment and other daily life scenes.However,with the spread of the new COVID-19 epidemic and the implementation of the normalized epidemic prevention policies,wearing masks has become a necessary choice for people to travel,which brings continuous challenges to the face image based authentication system.The mask covers nearly half of the face area,and only retains the image information of the eyes and forehead area.How to effectively use this part of the image data for accurate face recognition has become an urgent technical problem.This thesis proposes a face recognition method based on unoccluded areas,which solves the above problems by removing the interference of occlusion information and strengthening the extraction and utilization of feature information of unoccluded areas.The main works are as follow:1)A large-scale face recognition data-set with masks is constructed for model optimization training and effect testing.Based on two commonly used academic public datasets LFW and CASIA-Web Face,this thesis uses the Maskthe Face method to add masks to face images in data-sets,and constructs the "mask-wearing" face datasets Mask＿LFW and Mask＿CASIA-Web Face.2)An efficient and accurate image target detection network is selected as the basic network for face detection with masks.In this thesis,three representative target detection networks(Cascade RCNN,YOLOv5,Center Net)in three different network structures are selected for multi-scale comparative experimental analysis,and YOLOv5 with better comprehensive performance is selected as the basic detection network optimized in this thesis.3)A face detection network for partial occlusion is designed to improves the detection efficiency and accuracy.In this thesis,based on the YOLOv5 network,a multi-path feature extraction network is designed in the feature extraction module to improve the efficiency of the module for target retrieval;a lightweight residual module is constructed to accelerate the detection speed;the target classifier of the detector part is decoupled And position regressor to improve the positioning accuracy of the detection target.4)A face feature extraction network is designed to improve the recognition accuracy of partially occluded faces.This thesis integrates the idea of “attention mechanism” into the network design,strengthens the feature information of the eye area,improves the expression ability of the image features of the occluded face,and realize the image feature matching and retrieval functions of complete faces and partially occluded faces.The thesis compares the proposed network model with other commonly used face recognition models on the “masked face” data set constructed in this thesis.The experimental results show that,compared with the existing network,the network model proposed in this thesis has a certain improvement in detection robustness and recognition robustness for the face recognition problem of large-area occlusion by masks.