Face Recognition Method Research Based On Subspace Analysis

Face recognition is an important research field of biometric identification, facerecognition includes serval research areas, such as face detection, pre-process, featureextraction, classification and recognition, et al, in which, the feature extraction is one ofthe key step of face recognition, in present, there are many methods in this area, andsub-space based methods are become more and more popular for it’s good present-ability.In this paper, we mainly analyze the face feature extraction based on subspaceanalysis, at the same time, we also consider the effect of the distance metric in nearestneighbor classifier.The research mainly includes the following several aspects:Firstly, by reading a lot of domestic and foreign papers, this paper analyzes themethod of Principal component analysis (PCA) and Linear discriminant analysis (LDA),and summarizes the methods to slove the small sample size problem of LDA method.Secondly, we studied the Kernel discriminant analysis (KFD) method and Completekernel discriminant analysis (CKFD) method, for LDA method always meets the smallsample size problem and CKFD method did not unify the discriminant function, wepropose an improved CKFD method, in fact, the proposed method is equivalent to CKFDmethod in theory, but it can unify the CKFD’s two discriminant analysis, which is goodfor theory analysis.Thirdly, for the heavy computational complexity of one-dimensional methods, thispaper analyzes the two-dimensional principal component analysis (2DPCA) andtwo-dimensional linear discriminant analysis (2DLDA) method, proposes a new facerecognition method called block-weighted iterative2DLDA algorithm, the first step of thismethod is to divide the face image into three parts: the part above the eyebrows, the partbelow the nose, the part between the eyebrows and nose; then, extract the feature of eachblock using the iterative2DLDA algorithm.Fourthly, we studied the basic theory of the nearest neighbor classifier, for thedifferent features of the one-dimensional and two-dimensional method, we use differentnorm to classify.Experiments are carried out on ORL face database using the software of MATLAB,and the result is analyzed, the results of simulation experiments show the effectiveness andfeasibility of the proposed method.