Research On Face Recognition Model Based On Uncertainty Quantification

Face recognition is one of the important research fields of computer vision and object recognition,which is applied to daily life,security inspection.The advanced face recognition systems rely on the face embedding of deep neural network,which represents each face image as a deterministic point embedding.The distance between the embedded points can effectively measure the similarity between the corresponding face images.However,due to the low quality or incomplete input,the learned face features are low resolution,inaccurate or even nonexistent,resulting in false recognition.With the appearance of probabilistic face embedding,the idea of Uncertainty Quantification is introduced into the face recognition to improve the robustness.This dissertation realizes the following work:Firstly,ArcFace model is used as the baseline.The improved Res Net and Mobile Face Net are selected as the network framework to realize face recognition task.Each face image is learned into feature vector by ArcFace.According to the distance of face features,1:1 face verification and 1:N face identity query are realized.Through the experiment of ArcFace model,it is found that the recognition accuracy of ArcFace on unconstrained incomplete face images is low.Secondly,an improved method is proposed to improve the accuracy and robustness of ArcFace.Because probability embedding can capture the epistemic uncertainty,uncertainty is introduced into the learning of ArcFace.Face features learned by ArcFace based on Uncertain Quantization satisfy Multivariate normal distribution.The mean value represents the most likely face features,the variance represents the uncertainty degree of the learned features.In order to improve the robustness,face embedding is learned into a distribution with low uncertainty.Variance can also be used as an index of matching accuracy of feature space,which is important for risk control and identification.Finally,we compare the recognition accuracy and the recognition performance of incomplete face between the Uncertainty Quantization based ArcFace model and the original ArcFace model,and verify the effectiveness and robustness of the improved ArcFace model.