Face Verification With Deep-learning-based Feature Extraction

Driven by big data and deep learning,researches on automatic face verification have witnessed great progress in recent years.In this thesis,the deep-learning-based feature extraction algorithm applied to face identification and face verification task is extensively studied.To push automatic face verification system to real-time application,we propose a light-weight human face feature extraction framework based on deep learning theory,which shares both powerful feature representation and low complexity.The main work of this thesis are as follow:1.In this thesis,the research on face recognition and verification in the past ten years is reviewed,and the latest research background i.e.massive data and research method i.e.deep learning are pointed out.2.This thesis delivers a detailed introduction to the deep learning theory,including the history,char-acteristic,learning algorithm,etc.,and presents a tentative analysis on the underlying reasons for its strong nonlinear expressiveness.In addition,the 3 kinds of typical deep learning model,i.e.Deep Belief Network,Stacked AutoEncoder and Convolutional Neural Network are discussed and carefully analyzed.These work have important guiding significance for the design of face verification algorithms based on deep learning.3.In this thesis,a human face feature extraction framework based on deep learning model is proposed.The basic idea of this method is to perform a two-stage extraction pipeline.On the initial stage,we train mul-tiple convolutional auto-encoders with multi-scale face image patches to extract the low-level representation of the human face.Then,the extracted low-level features are fed into the immediately followed convolution-al neural network to extract the high-level semantic features.We argue that the framework is biologically plausible in that it accords with human brain's layered information processing mechanism.Furthermore,such scheme also reduces the model complexity,making it more viable for real-application.The simulation on Pubfig83 face identification task show that the proposed method does well in facial feature extraction,and has good performance for face identification task.4.In this thesis,we introduce a Joint Bayesian method for face verification,and deliver a detailed deriva-tion and discussion.For the limitation of the original algorithm(that either of the two components in the Joint Bayesian prior is zero-mean Gaussian random variable is a strong hypothesis),the method is further extended(to non-zero-mean hypothesis)so that it will be applied to a wider application range.Besides,a siamese-newtork-based deep learning model for face verification is also proposed and discussed.Simulation performed on both LFW View2 constrained and unconstrained protocols show the effectiveness of the both methods.