Integrating Two-Dimensional Morphing and Pose Estimation for Face Recognition

Abstract

Pose problem presents a challenge to face recognition methods because the shapes and features of the human face with large pose angle often appear quite different from the template. One approach to overcome this problem involves the estimation of the angle to which the face is rotated and then matches it to templates with the same pose. Another approach involves aligning the key facial features of the tested face to those of the template face before matching. However, these approaches can only achieve a moderate recognition rate, when the strict requirement of a low false alarm rate must be met. To accommodate the pose problem, this study integrates two-dimensional morphing with pose estimation. Two-dimensional morphing transforms various poses of a face image into a frontal view, thereby eliminating the center bias and pose variations and increasing the tolerance for pose estimation error. Experimental results show that two-dimensional morphing can significantly improve the performance of face recognition when dealing with severe pose variations. Combined with the pose estimation techniques, this integrated approach is capable of achieving very high recognition accuracy when the horizontal orientation is within +/- 45 degrees and the vertical orientation is within +/- 30 degrees. It also provides acceptably good performance with a horizontal orientation up to +/- 75 degrees and vertical orientation up to +/- 60 degrees.