Face recognition from 2D and 3D images using 3D Gabor filters

To recognize faces with different facial expressions or varying views from only one stored prototype per person is challenging. This paper presents such a system based on both 3D range data as well as the corresponding 2D gray-level facial images. The traditional 3D Gabor filter (3D TGF) is explored in the face recognition domain to extract expression-invariant features. To extract view-invariant features, a rotation-invariant 3D spherical Gabor filter (3D SGF) is proposed. Furthermore, a two-dimensional (2D) Gabor histogram is employed to represent the Gabor responses of the 3D SGF for solving the missing-point problem caused by self-occlusions under large rotation angles. The choice of 3D Gabor filter parameters for face recognition is examined as well. To match a given test face with each model face, the Least Trimmed Square Hausdorff Distance (LTS-HD) is employed to tackle the possible partial-matching problem. Experimental results based on our face database involving 80 persons have demonstrated that our approach outperforms the standard Eigenface approach and the approach using the 2D Gabor-wavelets representation.