| This study begins with an in-depth theoretic analysis and experimental evaluation of "state-of-the-art" face recognition techniques. We select three representative ones: eigenface approach, neural nets and elastic matching. Under a common statistical decision framework, these techniques were evaluated on four individual databases and a combined database with more than a hundred different subjects. Analysis and experimental results indicate that while considerable progress has been made, the effects of image formation such as the lighting and geometric transformations still remain problematic.; A solution invariant to the lighting variations is to match objects on real 3D surfaces that are independent from image formation. In the second part of this work, we examine the problem 3D reconstruction. The objective of this effort is to develop invariant 3D reconstruction, and therefore achieve robust recognition. However, 3D reconstruction is an ill-posed problem due to its implicit multiple-to-one mapping. The Karhonen-Loeve (K-L) expansion of a 3D surface proposed by Atick provides a solution for 3D reconstruction. While such a 3D reconstruction leads to lighting-invariant object recognition, it is still sensitive to image geometric transformations. We extend Atick's method to affine-invariant by modeling image geometric transformations as outcomes of a 3D affine transformation (scaling, rotation and translation), and present a new affine-invariant 3D reconstruction algorithm for object recognition. Experimental results on both synthetic and real 3D surfaces show that the performance of face recognition is improved by the affine-invariant 3D face reconstruction. |
