Research On Matrix-based LDA And2DLDA Algorithm For Face Recognition

Face recognition is a automatic technology according analysing the visualimformation of man’s face. In recent decenniums, it has developped quickly as thegreat process of computer’s calculating ability and supplys a good solution for theautomatic authentication. But the data of face image is big and it is easily influenced byillumination, angle and other natural factors, it is necessary to get a lot of imagesamples to march or train, and we need to reduce the complexity of the calculation andextract the key information for discrimination.2DLDA(Two Dimensional LinearDiscriminant Analysis) face recognition algorithm can reduce dimensions of faceimage data, It changes face image into two dimensional matrix and uses them for2DLDA. Compared with other methods, it is good at processing speed andidentification accuracy. This thesis mainly researchs how to improve2DLDA to get abetter accuracy. The main works in this thesis are as follows:(1) Both2DLDA and complex matrix LDA can fuse two different kindsinformation of face image. This thesis compares2DLDA and complex LDA on theeffort of face recognition with different number of samples for trainning by simulationexperiment, and analyses the reason of this result.(2) Considering the distance between target and average face image is always near,this thesis puts forward a new classifier called ED-AF on the base of euclidean distanceand average face, it can balance the efforts of nearest neighbour and average face tocorrect results of some test samples misrecognized because far from center. It has anew classifier proved to be better in accuracy by experiment in ORL face database.(3)2DLDA is good at data of faces which is near to each other in one sort, but ifthe data is distributed discretely,2DLDA may estimate a wrong projection matrix. Thisthesis researchs the definition of the within-class scatter matrixS wandbetween-classes scatter matrixS baccording to the principle of2DLDA, then it putsforward a new self-adaptive and enhanced method called E2DLDA, which rebuildsS wandS b, It is proved to make a better progress in face data which is compositedistributed by simulation experiment in ORL and FERET face database.