Research On Mask Occlusion Face Recognition Method Based On Deep Learning

Face recognition technology is one of the most important technologies in the field of biometrics.It uses a series of algorithms to extract features and compare and analyse face images for the purpose of identity verification and identification.However,face recognition algorithms do not always work positively due to real uncontrollable factors such as lighting,posture,resolution and occlusion.The presence of occlusion,among other things,can obscure most features on the face,resulting in a huge loss of information and a significant degradation in the performance of face recognition algorithms.In particular,mask occlusion under the normalization of epidemic prevention and control has a significant negative impact on facial feature extraction(e.g.mouth,nose,chin,etc.),prompting face recognition applications to consider mask occlusion problem.This paper addresses the problems of mask-obscured face recognition in unconstrained scenes and carries out research on mask-obscured face recognition methods based on deep learning.The main research work is as follows:(1)Aiming at the insufficient occlusion conditions in traditional face datasets,the small amount of data,the lack of occlusion in non-cooperative environments,and the insufficient demand for occluded face recognition research,large-scale mask occlusion datasets Z-CASIA and Mask-MS-Celeb-1M were produced.This paper uses the MTCNN model for face detection,and cleans the data to reduce the amount of sample redundancy,reduce the network computation.The dlib algorithm is used to delineate masked regions for faces and effectively create a large mask occlusion datasets.(2)To address the problem of low accuracy of traditional learning methods applied to mask-obscured face recognition tasks in unconstrained scenes,this paper proposes an attention mechanism-based face recognition method for mask-obscured faces.The method can be applied to both masked and unmasked recognition scenes.Two different attention modules are incorporated into the Inception-Resnet-v1 network structure to increase the weight of non-obscured areas and decrease the weight of obscured areas,so as to obtain more discriminative depth features of the face and improve the recognition effect.The experimental results show that the proposed algorithm can effectively improve the accuracy of masked face recognition and can accurately classify faces with masks.(3)To address the limitations of the method proposed in the previous chapter,which relies on CNN features and is prone to model overfitting and lack of interpretability,this paper proposes an EIRC-based face recognition method for mask-obscured faces.This method takes advantage of the strong interpretability and generalization ability of the capsule network,and fuses the improved model proposed in the previous chapter with the capsule network to change the internal structure of the network in order to overcome some of the limitations of traditional convolutional neural networks.The experiments show that the addition and improvement of the capsule network improves the robustness and generalization of the mask-obscured face recognition network compared with other advanced network models.(4)Based on the two aforementioned improved models,a mask-obscured face recognition system is designed and implemented.Based on the improved Inception-Resnet-v1 network model,a system framework and process for mask-obscured face recognition is designed.The relevant environment is set up for the system by establishing database,face detection and alignment,face feature extraction and face recognition processes.Finally,based on the analysis of the system requirements,modules for registration and login,information collection and face verification are developed,and functions such as detection and recognition of the system modules are introduced to complete the masked face recognition system.The improved algorithms proposed in the first two chapters are integrated into the system,and the visualisation effects of the two algorithms are compared with those of the benchmark model.The results show that the EIRC model works better compared with the benchmark model,verifying the effectiveness of the masked face recognition system and enabling the detection and recognition of faces under masking conditions.