Single Sample Face Recognition Under Complex Illumination Variations With Application To Driver Identification

With the rapid development of image technology,the application prospects of intelligent traffic monitoring system become more and more extensive.The driver face images extracted from the high-definition vichile images in intelligent traffic monitoring system can be used to recongnize with each other,or with the driver license face images and the suspect face images in known databases,which help to get the identification and location of the driver or suspect for the traffic law enforcement or criminal investigation.The driver face images are collected in outdoor environments,which are with complex illumination variations;and each vichile with one driver takes only one picture.Therefore,the driver face recognition problem is a typical single sample face recognition(SSFR)under complex illumination condidtions,which has a wide range of research value and application prospects in academic and commercial aspects.It is of great theoretical and practical value to study.This thesis concentrates on traffic driver face images,and studies two significant issues in SSFR:illumination treating and intra-class variation estimation.The main research results of this thesis are as follows:(1)This thesis proposes the high-frequeny facial features based on matrix decomposition to control the influences of local illumiantion variations,which often exist in driver face images.Firstly,inspired by the fact that the face image frequency model and the matrix decomposition model are with similar distribution characteristics,we construct the matrix decomposition model of the face image,which is the theoretical basis that the matrix decomposition algrothm can be used to extract the high-frequency facial features of the face image.Secondly,we develop the high-frequency facial feature model based on singular value decomposition(SVD)by using SVD bases with frequency interpretation and corresponding SVD weights.And then,the normalized SVD facial feature is proposed by designing a SVD weight normlization model,which can efficiently control the influences of illumination variations.Finally,we propose an adaptive SVD facial feature by constructing an illumination level estimation model based on SVD weights,which is more robust for various illumination variations,since its nonlinear parameter is adaptively selected according to the illumination level of the face image.Meanwhile,we propose an Orthogonal triangular with column pivoting(QRCP)facial feature by establishing QRCP bases with frequency interpretation and corresponding QRCP weights.Then a normalized QRCP facial feature is developed by normalizing QRCP weights,and an adaptive QRCP facial feature is constructed by using nonlinear QRCP weights to estimate the face illumiantion level.The experimental results show that the proposed methods perform much better than existing mainstream methods under various local illumiantion variations.(2)This thesis proposes an illumination invariant measure based on local regions to tackle the inflences of holistic illumiantion variations,which are normal for driver face images.Firstly,motivated by the fact that the face image in logarithm domain can be modeled by the additive model of the surface reflectance and illumiantion intensity,we construct an illumiantion invariant measure based on the local region in logarithm domain,which is more robust and less distortion than that measure in pixel domain.Secondly,since different local regions are with different discriminative powers,we design the discriminative weights and Gaussian weights,which are fused with several illumination invariant measures from different local regions to enhance robustness and weaken the distortion measure components caused by interference points.And then,we propose the illumination invariant measure facial feature based on local regions by introduced the sigmoidal function,which effectively tackles the high-frequency interference caused by partial fusion meausre distortion.Finally,the face image is divided into high-and low-frequency features by the matrix decompostion algorithm with frequency interpretation.An illumination preprocessing method is formulated by eliminating the illumination influences of the low-frequency features and correcting noise points of the high-frequency features.Further,we propose an illumination invariant measure facial feature based on frequency decomposition by above illumintion preprocessing method to robustly tackle severe illumiatnion variaions.The experimental results indicate that the proposed methods achieve higher recognition rates than existing mainstream methods under various holistic illumiantion variations.(3)This thesis proposes the adaptive approximation image reconstruction method to obtain high-quality approximation images,which cope with lack of face intra-class variation informtion in SSFR.Firstly,as the rank of a face image matrix describes its sparsity and image size variation,the relationship model between the face image rank and the approxiamtion image is constructed,and an adaptive approximation image recostrution method based on SVD is developed to make SVD based approxiamtion image reconstruct adaptively.Then,the fast optimial projection space of the Direct linear discriminant analysis(DLDA)algorithm is formulated by utilizing SVD and QRCP to diagonalize both intra-and inter-class scatter matrices of DLDA simultaneously.Secondly,an adaptive approximation image reconstruction method based on LU decomposition is proposed.This method utilizes Lower-upper(LU)decompostion algorithm to decompose the face image and its transpose into two independent decomposition image sets,and defines the criterion of establishing the standard decomposition image set.An energy function based approxiamtion image reconstruction model is designed,and the high-quality approximation image can be attained by the optimal value of the energy function esimation model.The experimental results illustrate that the proposed methods outperform existing approximation image reconstruction methods for SSFR.(4)Face intra-class variation estimated by the auxiliary set is very important to improve the performance of SSFR.This thesis proposes the illumination invariant feature based sparse representation method to tackle two important issues of SSFR:auxiliary face intrinsic information elimination and face intra-class variation estimation.Firstly,since the additive model of the facial feature decomposition is helpful to eliminate auxiliary face intrinsic information,we design two additive models of facial feature decomposition(high-and low-frequency additive(H&L)model,reflectance and illumination additive(R&L)model),which are introduced into the sparse representation based SSFR model.Then,the sparse representation model based on facial feature decomposition can be established,which separates the mutual inflence of illumination and face intrinsic information,and eliminates auxiliary face intrinsic information.Secondly,the illumination invariant feature based sparse representation model is proposed by the fusion of illumiantion invariant feature and logarithm image,which insulates illumination interference,eliminates auxiliary face intrinsic information,and preserves whole discriminative information of the face image.Finally,a vector projection classification algorithm is develped by considering the impact of image vector class center to classification result,which is introduced into the illumination invariant feature based sparse representation model with approximation images.Then,an illumination invariant feature sparse representation vector projection classification method based on approximation image is proposed,which achieves illumination interference insulation?auxiliary face intrinsic information elimination?efficient face intra-class variation estimation and discriminative information of the face image maintenance,and maximumly improves the performance of SSFR under complex illumination variations.The experimental results indicate that the proposed methods are much superior to existing mainstream methods for SSFR under complex illumiantion variations in terms of recogniton rate.