Face Detection Based On K-L Transform

Face Recognition is a important method in personal identification field .which can be used wildly in some fields such as certificate verifying, Public Searching-Escaping , credit card verifying, ATM. comparing to other personal identification methods such as fingerprint, hand, retina, and iris, Face Recognition is more directly, more friendly and more convenient.A complete Human Face Recognition System should include Human Face Detection, Feature Extraction, and Match Recognition. Face Detection is the first step among them and is also the important step in Face Recognition. This paper is for the research of face automatic Recognition Using K-L algorithm and Singular Value Decomposition after a few Illumination Compensations. We bring up a new engineering method. It can automatically choose the number of eigenface that match appointed "detect precision" , then exclude the axises with lower energy in eigenface space and make it easier to find the energy centralized axis in eigenface space. Firstly, extracting feature basing on the feature of K-L outspread formula. Using average sample of each one and average sample of all people among input samples, we form between class scatter matrix and make it as producing matrix of K-L transform. We do Singular Value Decomposition with this matrix and then we can gain eigenface space. At the same time, we can also gain a set of projection coeffient which is the algebraic feature of the training samples. Secondly, we project the picture to the space and gain its algebraic feature, thencontrast the algebraic feature of the picture and training sample. During the contrast, we choose the number of algebraic feature of training samples that is corresponding to the number of eigenface basing on appointed precision. We then calculate Euclidean distance with the algebraic feature of the picture and the selected algebraic feature of training samples and tell if the picture is face in terms of this distance.From the result of this experiment we can know that this methord is robust. It can adapt for the environment with the number of training sample changing sharply. It can effectively exclude feature axis with lower energy in eigenface space. When the number of training samples is big, it can effectively lessen the work of choice of energy centralized axis in eigenface space and then gain least eigenface subspace that is mostly like face.