Research On Unconstrained Face Recognition Based On Convolutional Neural Network

Convolutional Neural Network(CNN)is a new kind of Artificial Neural Network(ANN)which is combined by the method of ANN and deep learning.Convolutional Neural Network has been widely used in Computer Vision because of its exceptional performance in the field of image processing.Unconstrained face recognition is a difficult task for traditional face recognition method.While CNN has been proven to reach a satisfactory result when deal with this kind of problem.The purpose of this thesis is to propose a deep convolutional neural network for unconstrained face recognition.This thesis mainly include: The summary of the basic theory and the introduction of the basic structure of convolutional neural network.Taking the classical network LeNet-5 as an example,the general structure of the general convolutional neural network model is also illustrated.The structural features and parameter configuration of VGG-19 which have achieved good results in face recognition have been analyzed in this thesis.After finding the reasons for its high hardware requirements and training difficulty,an improved method for VGG-19 network model is designed.The fully connect layer of VGG-19 has been changed into a new architecture,by reducing the training parameters of the network and replacing the random initialization method by the parameter initialization method,the hardware requirements have been reduced,and the computing efficiency has been improved with the ensurance of high accuracy.The new designed convolutional neural network model is trained and tested on the authoritative face databases FaceScrub has achieved a high recognition accuracy.The test result of this method has been compared with those of traditional face recognition methods as well as other two methods based on convolutional neural network.Finally,the advantages and shortcomings of this proposed convolutional neural network are summarized.