3D Face Recognitionacross Pose Extremities

Abstract

In this paper, a mathematical model for 3D face image registration has been proposed with poses varying from 0 to ± 90˚ across yaw,pitch and roll. The method which has been proposed in this paper consists of two major steps:- 3D face registration and comparison of the registered image with any neutral frontal posed model in order to measure the accuracy of registration followed by recognition. In this 3D registration and recognition model, a 3D image is transformed from any different pose to frontal pose. After applying the algorithm on the Bosphorus databases, our proposed method registers the images with poses ranging from 0 to 20° with an average rotational error ranging between 0.003 to 0.009. For poses with an orientation of 40° to 45˚, the average rotational error was 0.003 to 0.009 and for poses with 90° the average rotational error was 0.004. Features are extracted from the registered images in the form of face normals. The experimental results which were obtained, on the registered facial images, from the 3D Bosphorus face database, illustrate that our registration scheme has attained a recognition accuracy of 93.33%.