Research And Application Of Face Recognition Technology Under Mask Occlusion

With the continuous development of deep learning,the accuracy and stability of face recognition has been significantly improved.Traditional face recognition algorithms rely on the acquisition and matching of facial features,such as eyes,nose,and mouth.However,with the outbreak of the novel coronavirus pneumonia(COVID-19),wearing masks has become a necessary measure in public places.As a result of wearing masks,most of the facial features are obscured,which leads to a significant decrease in the accuracy of traditional face recognition algorithms.In order to improve the accuracy of face recognition under mask occlusion,thesis proposes a scheme to repair the missing face features in the mask occlusion area first and then perform face recognition.The main contributions of thesis are as follows:Facing the situation that key features of faces are significantly reduced after wearing masks,thesis proposes an improved GAN-based face restoration algorithm to repair the missing face features while removing mask occlusion and with realistic effects.The face restoration algorithm model consists of a generator and a discriminator.The backbone of the generator uses a U-Net network that can efficiently extract high-level semantic features from face images,and then introduces residual blocks and attention units to optimize the generator.The discriminator combines a global face discriminator and a local face discriminator for the mask region,while combining various loss functions such as perceptual loss and reconstruction loss for adversarial training.Finally,the restoration effectiveness of the algorithm is evaluated qualitatively and quantitatively on the FFHQmask mask masked face dataset.The experimental results show that the algorithm has higher similarity and more realistic visual effects.In order to more reasonably and effectively extract the real face features in the face and the face features generated by the restoration,thesis proposes an improved Mobile Net V3-based face recognition algorithm.The algorithm optimizes the feature extraction network and attention mechanism on the Mobile Net V3 model with both accuracy and speed,and can focus on the real face features and reduce the weight of the features in the maskobscured area when performing face recognition under mask obscuration,so as to reduce the influence of mask obscuration and background interference on face recognition.The experiments show that the improved Mobile Net V3-based face recognition algorithm alone can improve the accuracy rate by 5.48% compared with other better face recognition algorithms.And combined with the use of the improved GAN-based face restoration algorithm proposed in thesis,it can improve the accuracy by another 1.12%.Also the algorithm has better performance in practical scenes.In addition,based on the improved GAN face repair algorithm and the improved Mobile Net V3-based face recognition algorithm,thesis designs a face recognition management system for wearing masks to realize visual management,which has better performance in practical applications.