Gender Determination From Single Facial Image By Utilizing Surface Shape Information

This paper investigates the feasibility and the effectiveness of performing gender determination from single 2D facial intensity image for each subject by utilizing recovered 3D facial surface shape information. The 3D facial surface shape in the form of needle-map, the fields of facial surface normals, is recovered from single 2D facial intensity image by using principal-geodesics-based shape-from-shading technology (PGSFS). An important fact is that the original 2D facial image is implicitly encoded into the 3rd component of the recovered 3D needle-map. The recovered needle-maps lying on high-dimensional Riemannian manifold are projected onto a lower-dimensional sub-manifold space by applying a special manifold learning technology - Principal Geodesics Analysis. Liner discriminant analysis is then performed on the projected needle-maps to find the optimal discriminant direction(s) for gender determination. Experimental results on FERET database demonstrate that the proposed method, PGA+LDA on the recovered 3D needle-maps, outperforms the classical method, PCA+LDA on 2D facial intensity images.