Chinese Journal of Computers   Full Text
Title	Pose and Illumination Invariant Face Recognition Based on HMM with One Sample Per Person
Authors	HU Feng-Song1) ZHANG Mao-Jun2) ZOU Bei-Ji3) MA Jun-Rong1)
Address	1) (College of Computer and Communication, Hunan University, Changsha 410082) 2)(College of Information System and Management, National University of Defense Technology, Changsha 410001) 3)(College of Information Science and Engineering, Central South University, Changsha 410083)
Year	2009
Issue	No.7(1424—1433)
Abstract &   Background	Abstract In this paper a novel pose and illumination invariant face recognition algorithm that based on HMM with one sample per person is proposed. Firstly, by learning the train sets that has been fitted with Candide3 model, fitting algorithm automatically fitted the frontal facial input image with Candide3 to get the 3D shape. Then the specifically 3D face model is reconstructed by synthesizing texture to the fitted 3D shape. The new images under different pose can be generated by transform the 3D face model. Decompose the new images to 9 harmonic images’ linear combination. By changing the 9 coefficients the new images under different illumination condition can be generated. All of these new images under different pose and different illumination condition make up the train set of individual face hmm. Experimental results show that this method can effectively avoid the recognition rate bring down caused by the pose and illumination normalization is not effective sometimes and can be better fitting the pose and illumination invariant face recognition. Keywords 3D face reconstruction; harmonic image; hidden Markov model; face recognition Background This work is accomplished by the union of face recognition researchers from Hunan University, Central South University and National University of Defense Technology. This work is partly supported by the National Natural Science Foundation of China under project “Trace and Analyses Method of Body Empty Out Roll Movement based on Video” with grant No.60673093， National Natural Science Foundation of China under project “Key Technology Research on Major Engineering Disaster oriented Numerical Value Integrated Simulate Plat” with grant No.90715043 and Natural Science Foundation of Hunan Province in China under project “Research on Video Analyses Method and Its Application of Body’s Bounce Movement” with grant No.07JJ3125. The group has been concentrated in Patter recognition, image processing and image recognition for decades. In the past several years, plenty of research work of the group has been done on the basic theories in face recognition，as well as the practical engineering techniques. There were several papers published or accepted by some well known journals or proceedings，and many of the research fruits have been successfully applied to practical applications and are going to be industrialized. This paper solved the problem of face recognition under variant illumination and poses effectively; the face recognition rate is 20% higher than the recognition rate of the four images in the newest paper in the research region. The theory base of this paper is individual 3D face generation method proposed by authors, which is published in Journal of Hunan University. This paper make use of Candide3 model, combining an importation of frontal face image samples to produce individual person’s face 3D model, thus obtaining the variety of poses. Experimental results show that this method can effectively avoid the recognition rate bring down caused by the possible inefficient pose and illumination normalization, and can be better fitting the pose and illumination invariant face recognition.