The Research Of Face Recognition Method Wearing Mask Based On Instance Segmentation

At present,the performance of face recognition technology based on deep learning has made continuous breakthroughs and is becoming increasingly saturated.Its development status cannot meet the needs of the new situation of social epidemic.The new crown virus swept the world,China has entered the normalization of new measures and control stage.Wearing masks to travel is still a supply and demand measure of transportation hubs.The large-scale occlusion of the face to the traditional face detection technology accuracy has dropped significantly,similar masks,glasses and other occlusions seriously interfere with the recognition results.Face recognition system work is difficult to carry out.The positioning of the occlusion area of the face wearing the mask and the extraction of the key features of the eyebrow are the key elements to improve the accuracy of face recognition with the face wearing the mask.This project considers the recognition accuracy and model generalization requirements,and studies the modeling and recognition model of image depth feature enhancement under large-area occlusion of human faces.Focusing on solving the problem of low recognition accuracy caused by the failure of face detection and insufficient effective features in wearing masks.The paper jumps out of the traditional face detection method based on rectangular box positioning,and proposes a face recognition method for maskwearing masks combined with instance segmentation.The specific research content is as follows:Aiming at the problem of the failure of face detection method caused by large-area occlusion.An alternative method of segmentation method for mask-wearing face instances under multi-stage edge segmentation is proposed.Firstly,the idea of group attention is added to the backbone extraction network of Mask Region-based Convolutional Neural Network(Mask R-CNN),which enhances the connection between features of different sizes and improves the generalization ability of target segmentation of different sizes.In addition,a new intersection ratio loss function is designed,which is fed back to the prediction box from three aspects: the overlap,the center point distance and the size.So that the prediction box converges to the target area more accurately.It generates a coarse segmentation mask and the corresponding target frame.Aiming at the problem that Mask R-CNN segmentation boundary accuracy is low and cannot provide sufficient effective area features for face recognition.A multi-stage edge segmentation method is proposed.A linear logical relationship between occluded and unoccluded faces is established and stored in the mask occlusion dictionary,based on the feature masks of occlusion features and face edge occlusion areas.Mutational pixels are extracted to reassign edge features,the occlusion classifier is trained by bilinear differential twin network(Res Ne St-Pairwise Differential Siamese Network,R-PDSN).Combines the multi-stage iterative optimization method to resegment the coarse segmentation mask,continuously optimizes the target edge information,and gradually realizes fine segmentation.Aiming at the problem of low accuracy of face recognition with masks.A network model based on eyebrow attention mechanism and a training method based on decision-making fusion are proposed to build the discriminating features of two forms: complete face and occluded face.Its enhance the recognition ability and adaptability of the model to random faces,and improve the accuracy of face recognition.Based on the above methods,the improved Mask R-CNN and R-PDSN are used to segment the eyebrow area features of the face wearing the mask.The feature extraction network is used to focus on this location to complete the face recognition task.The public mask wearing face dataset is used for training and testing,and the average pixel accuracy of the Mask R-CNN of the benchmark model is improved by2.68%.The target detection accuracy is more than 98% in terms of segmentation and detection accuracy.In terms of recognition,the proposed method is improved by 15.70%compared with the traditional face recognition method.Under this model,a face recognition system wearing a mask is designed,which has good real-time performance.The recognition efficiency of the system is significantly higher than that of the popular face recognition technology.The system provides new ideas for the landing of face recognition wearing masks.It is convenient to transplant to identity recognition systems such as face ID verification and security monitoring,which has certain research significance and application value.