Local Preserving Projection Based Face Recognition Under Uncontrolled Environment

Currently, face recognition has a promising application in the fields of securityprotection, video surveillance and identity authentication, which is regarded as an activeresearch issue in pattern recognition and computer vision. It still remains to be quitechallenging to conduct automatic face recognition successfully that fulfill the demands forpractical application under uncontrolled environment with variant illumination and poses.Therefore, this thesis focuses on the research of LPP-based face recognition algorithmwhich is robust to illumination and poses variation on the basis of the existing facerecognition researches.Feature extraction is the most important procedure in face recognition, here SODLPPfeature extraction algorithm is proposed by improving the ODLPP algorithm concerningthe selection of projection eigenvectors. In order to solve the illumination and posesproblem, we use multi-scale self quotient image as the illumination preprocessing methodto eliminate the light spots and shades on the facial images partly, then a multipoint imagewarping technique is employed to generate multi-poses facial images from frontal facialimages for training. At last we combine Gabor wavelet transform with LTP operator toextract LGTP features to represent facial images for face recognition which can overcomethe negative effects caused by illumination, poses and facial expressions variation.Experimental results on the face databases ORL and YALE B show that the proposedSODLPP algorithm has better performance compared with ODLPP algorithm. Meanwhile,we can draw a conclusion from the related experiments that the robustness of our facerecognition algorithm against illumination and poses variation can be enhancedsignificantly by employing the proposed methods of illumination preprocessing,multi-poses facial images generation and LGTP based facial feature representation.