3-D Face Analysis and Identification Based on Statistical Shape Modelling

Abstract

This paper presents an effective method of statistical shape representation for automatic face analysis and identification in 3-D. The method combines statistical shape modelling techniques and the non-rigid deformation matching scheme. This work is distinguished by three key contributions. The first is the introduction of a new 3-D shape registration method using hierarchical landmark detection and multilevel B-spline warping technique, which allows accurate dense correspondence search for statistical model construction. The second is the shape representation approach, based on Laplacian Eigenmap, which provides a nonlinear submanifold that links underlying structure of facial data. The third contribution is a hybrid method for matching the statistical model and test dataset which controls the levels of the model's deformation at different matching stages and so increases chance of the successful matching. The proposed method is tested on the public database, BU-3DFE. Results indicate that it can achieve extremely high verification rates in a series of tests, thus providing real-world practicality.