Face Transformation And Recognition From Near Infrared To Visible Lighting Images

Face recognition is a hot topic in the fields of pattern recognition and computer vi-sion. Recent years have witnessed many successful applications of face recognition inkinds of situations. In practice, most of face images are captured under visible lightingand thus complex and varied visible lighting conditions greatly in?uence the perfor-mance of face recognition. Lighting invariant face representation is always one of thekey issues of face recognition.Near infrared imaging technique has been applied to accomplish the lighting in-variant face recognition to some extent. In near infrared face recognition system, anatural requirement is that both enrolled face images and query face images should becaptured under near infrared lighting. However, in many important situations whereface recognition works, face images can only be captured under visible lighting, suchas video surveillance, photo-based identification. Therefore, a new problem emergesthat of cross recognition between near infrared and visible lighting face images. Dueto their different imaging methods, near infrared and visible lighting face images fromsame person can be significantly different in appearance. But in the view of humancognition, they still can be recognized as same person. That means two things: 1)thereexists some underlying correspondences between near infrared and visible lighting faceimages and 2)some invariant features can be extracted between them. From above twopoints of view, this paper presents two kinds of methods to achieve the face transfor-mation and recognition from near infrared to visible lighting images.Main contributions are as follows:Manifold mapping based face transformation. A hypothesis of continuous map-ping is introduced to build the correspondence between local appearance mani-folds of near infrared and visible lighting faces. Based on this hypothesis, Bothimplicit and explicit mapping can be learned from training database consistingof simultaneously captured near infrared and visible lighting face images. Soface transformation can be achieved by these mappings. In transformation ex-periments, Peak Signal Noise Ratio(PSNR) is used as criterion to measure thequality of face transformation. Finally, recognition experiments are conduced on synthesized and real visible lighting face images.Invariant features based face recognition. Firstly, invariant features between nearinfrared and visible lighting are explored through the analysis of imaging model.Three local patterns are proposed to extract histogram features so as to form facerepresentation. Further, discriminative features are selected by Fisher criterion.Finally, face recognition are accomplished by matching invariant features. Ex-perimental results show the effectiveness of proposed methods.At the end of this paper, a summary is presented and some future works are indi-cated.