Face Recognition Based On High-dimensional Feature And Deep Learning

With the advent of the information age,people are paying more and more attention to information security.As a biometric identification technology,face recognition that cannot be reproduced or easily be imitated is of high security.Therefore,it is widely used in authentication,visual communication,intelligent monitoring,daily security,finance,e-government,public file management and other fields.Generally,face recognition is to judge by computer whether two facial images are from the same people or who is in the image based on the information in facial images.Therefore,it is important to extract effective features of facial images.This thesis studies on extracting of features and recognizing of facial images.Firstly,this thesis improved the algorithm of constructing high dimensional feature by Chen Dong et al.In the paper of Chen Dong et al.,facial images are zoomed into several different dimensions.In each dimensional image,a fixed size of block centered at each key point is chosen and then divided into sixteen units.Feature vector is extracted for each unit.After features of all units are extracted,they are combined into the feature of a block.In the end,all features from all blocks in all dimensional images are combined into features of a facial image.In this thesis,one unit centered at each key point is added into each block,the improved high dimensional feature is presented.Finally,experiments are conducted on LFW face dataset to compare the high-dimensional LBP,Gabor,HOG,SIFT features and the corresponding improved high-dimensional features.Experimental results show that the improved high-dimensional feature is superior to the high dimensional features proposed by Chen Dong et al in recognition rate.Secondly,the improved method for constructing high-dimensional features is applied to LPQ features,the concept and construction method of high-dimensional LPQ feature are proposed.On the basis of this,a new face recognition algorithm based on high-dimensional LPQ feature and joint Bayesian algorithm is proposed.Similarly to the improved high-dimensional feature,the LPQ feature is extracted from each cell.Then the unit LPQ features from all blocks in all scales are combined to obtain improved LPQ features.At last,PCA is used to reduce the dimensionality of LPQ,and Bayesian is used to recognize.Experiments are conducted on the LFW face database with features of LBP,HOG,Gabor,SIFT,and their corresponding high-dimensional features.The results show that recognition rate of our proposed high-dimensional LPQ feature is higher than other features and the corresponding high-dimensional features.Thirdly,this thesis studies on whether dimension reduction with PCA method is helpful for improvement of face recognition rate when the feature vector is extracted by the convolutional neural network.The deep feature of the image is extracted by using the VGG model,then the dimension is reduced by the principal component analysis(PCA),lastly,cosine classifier is used to recognition.Experiments on the LFW face database show that the recognition rate can be improved by using the PCA to reduce the dimension of the deep feature obtained by the VGG network.