Research On Deep Learning Based Feature Augmentation And Metric Learning Algorithm For Face Recognition

Face recognition has always been a research hot spot in the field of artificial intelligence.Compared with other biometric authentication methods,face recognition has broad application scenarios in attendance,payment,station entry,boarding,and monitoring of specific objects.The current mainstream algorithms are mainly deep learning methods based on convolutional neural networks.Existing methods require the face of the subject to have good lighting conditions,the expression and posture cannot be changed too much,and there can be no occlusion.In a constrained environment,face recognition based on deep learning has surpassed the recognition ability of human beings.However,under unconstrained conditions,such as the video recorded by a surveillance camera,the subject has a side face,or is blocked,wears glasses or a mask,blur,low resolution,exaggerated expression,large posture changes,large lighting changes,etc.,under these changes,the latest recognition models and algorithms cannot obtain highly discriminative features.Besides,because face detection and recognition are two different models and algorithms,the existing algorithm does not consider the inherent relationship between the two.If the detection result is unfavorable,the recognition result is also unfavorable,resulting in adverse robustness of the existing face recognition algorithm.In addition,most of the current face recognition systems are based on very deep large-scale networks and do not support operation in embedded systems.Therefore,it is important to study face detection and recognition methods based on lightweight networks and design highly discriminative feature representations.In response to the demand for face recognition algorithms in unconstrained environments,this dissertation focuses on key technologies such as deep learning-based feature representation,feature fusion and enhancement,and metric learning based on existing face detection and recognition algorithms.Based on the convolutional neural network,a lightweight face recognition pipeline is designed and implemented,including a multi-scale face detector and a template-based face recognizer.The experiment results show that the lightweight Res Net-18 model has high accuracy,and the average accuracy,false recognition rate,and rejection rate on the IJB-C data set are close to the backbone benchmark network Se Net50,meeting the realtime and robustness requirements of future face recognition systems.The main research contents and innovations of the dissertation are as follows:(1)A new face detection method with dense Anchor is proposed.By studying the problems of single network anchor box matching,this dissertation adds auxiliary networks,including auxiliary loss function items,and finally increases the matching probability between the preset anchor box and the ground truth face box.Through experiments on several public face detection benchmarks,the effectiveness of the proposed detection method is verified.The method detects 892 faces on the world's largest selfie;(2)A method of fully fusing high-level strong semantic features and shallow high-resolution features of Convolutional Neural Network(CNN)is proposed.The existing feature fusion methods directly fuse the feature maps of different layers of CNN.There are redundant and abnormal feature values,which cannot guarantee the fusion of complementary and diverse features.Therefore,fused features may not be useful for detection and recognition.Based on the feature complementarity and diversity of heterogeneous networks,this dissertation proposes a dynamic feature augmentation algorithm for feature maps,which can be easily integrated into existing CNNs.The feature augmentation pyramid generated by this method improves the effective representation and extraction capabilities of face features in unconstrained environments.When TAR@FAR=0.1,the accuracy of face verification on IJB-C is increased by 16%;(3)A metric learning method based on KL divergence is proposed.In the template-based face verification problem,the traditional method is to use a set of features to represent the video or template,and each feature corresponds to its constituent image or frame.This method has high complexity in calculating the similarity of two frames of video,large memory consumption,and cannot be expanded with a large number of videos.The metric function in this dissertation includes two components: fidelity constraint and similarity constraint.The fidelity constraint condition calculates the distance between the newly learned feature distribution and the original feature distribution so that the newly learned feature distribution is close to the original feature distribution.The similarity constraint ensures that the similarity of the same template is greater than the similarity of different templates.According to the scores of the earlier face detection pipeline,the face sent to the face recognition system is dynamically adjusted.This method is verified on IJBC,and finally,the lightweight model in this dissertation can effectively perform face recognition on IJB-C,and the accuracy is increased by 46% when TPIR@FPIR=0.01;(4)An improved the loss function of the existing single shot multibox detector(SSD)detection model is proposed to predict face pose.This method makes full use of SSD's inherent classification and regression capabilities and avoids the high coupling and time-consuming shortcomings of existing face pose prediction methods.Using the bin method,the continuous angle of the face deflection pose is converted into multiple specific categories required for training,the 3D angle regression problem is converted into the angle classification and face box regression problem,and the model directly outputs the Euler angle(yaw angle,pitch angle,and roll angle).The average mean precision predicted by this method of AFLW2000 and 300 WLP is 6.01° and 2.38°,respectively.This dissertation studies face recognition based on deep learning.Based on the existing Arc Face,the linear function is used instead of the cosine function,which avoids the limitation of Arc Face for small models to start training with Soft Max.Using the feature fusion and augmentation algorithm proposed in this dissertation,and the depth metric learning method based on KL divergence,for the new benchmark data set IJB-C released by NIST,under the unconstrained environment with full pose and illumination variation,the recognition rate of rank-1 based on Res Net-18 is improved from 26% to 68%.