Research And Application On Feature Extraction Technique For Face Recognition

Face recognition is one of the hot topics in pattern recognition and computer vision.As an important biometric identification technology,face recognition has been widely applied in daily life,such as social security,surveillance,authentication etc..The direct aim of face recognition is to identify and verify the individuals.Face recognition in real scene meets a great challenge due to the variants of illumination,occlusion,pose and expression etc..A typical face recognition system usually consists of five phases:face detection,face preprocessing,feature generation,feature extraction/selection and matching&identification,in which feature generation and extraction have the important effect on the accuracy of face recognition.Here,main attentions have been paid to feature generation and extraction phase to extract effective facial features from original images and to extract useful identification information from high dimensional facial features,so as to improve the performance of face recognition.In this manuscript,typical local feature descriptor,sparse theory and low rank representation as well as collaborative representation have been well studied and abundant experiments have been conducted on public face databases to verify the proposed methods.The highlights are presented as follows:(1)For some feature descriptors such as Local Ternary Patterns(LTP),appropriate threshold shall be selected in feature generation in order to overcome different noise problems.Inspirited by Weber's law,Adaptive Local Ternary Patterns(ALTP)and Central Symmetric Adaptive Local Ternary Patterns(CS-ALTP)are proposed,which will automatically choose corresponding threshold according to the gray value of pixel to solve the defect where fixed threshold cannot adapt to the change of pixel.In addition,CS-ALTP has lower feature dimensions than ALTP.Proved by ORL&FERET face database experiments,ALTP and CS-ALTP achieve a better performance than traditional feature descriptors do in recognition rate.(2)Low rank representataion required the dictionary to be complete,so dimension reduction is still an important work.First,a low rank representation based projections(LRRP)method is presented for face recognition.In LRRP,the nuclear graph are constructed by using the low rank representation theory,the local compactness information and the total seprateness information of the samples are characterized on the basis of the method.Second,utilizing low rank representation in the reduced subspace to build the affine graph,a two Steps Iterative Low Rank Representation based Projections(TSILRRP)method is further offered.Third,based on sparse representation based classification strategy,Low Rank Representation Analysis(LRRA)is proposed to feature extraction.In LRRA,it does not need to learn the affine graph in graph embedding learning.Experiments on the FERET,AR,ORL face database and PolyU KFP database show the effectiveness of the proposed methods.(3)Based on that collaborative representation based classification(CRC)has good performance and is fast,a Collaborative Representation based Projections(CRP)method is proposed for feature extraction,which enriches graph embedding learning framework.CRP uses L2 norm to build the affine graph to characterize the local compactness and total seprateness of the samples.CRP constructs the objective fuction under the Fisher discriminant analysis framework,it is solved using generalized eigenvalue decomposition.Further,a Kernel CRC(KCRC)method is put forward to strengthen the robustness of CRC.Kernel CRC utilizes kernel techniques to map the original data into high dimensional space to optimize.Experimental results on several public databases show that the proposed methods surpass the classical method.(4)To identify low-resolution face image from video surveillance,high-resolution images and low-resolution images are regarded as two different groups of variables,and projected to the same subspace via canonical correlation analysis(CCA)to realize the matching of images from different dimensions.By this method,a low-resolution degradation face recognition method is proposed to overcome the problems of inconsistent and low resolution.The experimental results on Extend Yale B,ORL and AR face database demeonstrate that the proposed method is robust to low-resolution image.