Recently, face recognition technology is being paid more attention not only in the academe but also in the industry. How to improve face recognition performance is of significance in terms of theory and application. Illumination and the change of human face's pose have a great impact on face recognition. The former can be solved by using near infrared image, and the latter can be improved by face pose estimation. Therefore, the estimation of human face's pose is the key subject in face recognition.Face pose estimation is an image processing that estimate the pose of the human face in three-dimensional space using a two-dimensional image. Face pose rotation consists of three degrees:rotation around yaw. tilt and rotating in the plane. This paper, beginning with how to estimate the pose of these three degrees, mainly solves the following key problems:(1) the precise location of the feature points (including the eye point,nostril.the nasal tip and the mouth center-line):using the edge information to fit an ellipse to locate the pupils, using Quoit filter to locate nostrils, using the highlighted features and the relationships with nostril to locate nasal tip. using the characters that mouth centerline is a dark line with normal expression to locate mouth center line;(2) the face pose modeling:assuming face is a space cylinder, we find the mapping from three-dimensional world coordinate system to the two-dimensional image plane coordinate system, and theoretically deduce the relationship between the posture parameters in the three-dimensional space and the organs characteristic position in two dimensional images. Based on the relationship functions and the precise locating of the facial feature points, high-speed and accurate estimation of the face pose parameters can be made.This paper also carries out a large number of experiments to prove the accuracy of feature location via this method, the average feature location error in the near-infrared light image library is around three pixels. We also verified the accuracy of head pose model under a large number of face pose library, the average estimation error is about5degrees, it proves this model achieved high accuracy estimation results. Finally, we implemented a face pose automatic estimating system which can automatically determines the angle of rotation on the near-infrared light images with good performance:the average estimation error is less than5°, and the speed of processing a single image is less than20ms. |