Research And Implementation Of Age Invariant Face Recognition Based On Deep Learning

Face recognition is one of the most widely used and important fields in computer vision.Security monitoring,attendance,remote user authentication,financial payment and even social entertainment have already spawned related applications.The research of face recognition has lasted for decades,from the earliest eigenface algorithm to the current convolutional neural network algorithm.Thanks to the improved hardware computing ability,big data support and deep learning algorithms,computer vision technology including face recognition as the representative entered the people's vision again.Although the related fields such as face recognition under different expressions and gestures have made good progress,age invariant face recognition,or cross age face recognition,is still a challenge.Today,the field of cross age face recognition has become increasingly important,and has a wide range of applications,such as the search for missing children,identify criminals and passport authentication.In this thesis,based on the related theories and methods of deep learning,a new neural network model called coupled auto-encoder networks is proposed to handle age-invariant face recognition problem.The main work and innovation are as follow:1.Auto-encoder is used in deep learning to model.As a typical unsupervised learning method,auto-encoder can learn hidden representations automatically from inputs(like images).2.Observed that age variation is a nonlinear but smooth transform.Two shallow neural networks are used to fit complex nonlinear aging and de-aging process because a single-hidden-layer neural network can fit any complex smooth function.3.Combined with the above two points,coupled auto-encoder networks(CAN)is proposed,which is a couple of two auto-encoders which bridged by two single hidden layer neural networks.4.Based on CAN,a nonlinear factor analysis method is proposed to nonlinearly decompose one given face representation into three components which are identity feature,age feature and noise,where identity feature is age-invariant and can be used for face recognition.5.A two-step learning algorithm is proposed to train CAN to separate identity feature for age invariant face recognition.6.Experiments on three public available face aging datasets: FGNET,CACD and CACD-VS show the effectiveness of the proposed approach.