Facial modelling for forensic facial reconstruction and identification
Dirk Vandermeulen (ESAT/PSI K.U.Leuven), Peter Claes (University of Melbourne), Sven De Greef (Forensic Dentistry, K.U.Leuven), Guy Willems (Forensic Dentistry, K.U.Leuven), Paul Suetens (ESAT/PSI K.U.Leuven)
We present two automated methods for estimation of the facial outlook from human skeletal skull remains. The first method uses a database of a little less than 400 3D facial surfaces, augmented with soft tissue thicknesses, measured with ultrasound at 52 anatomical landmarks. A standard PCA-model (including also extra individual characteristics such as age, sex and BMI) is constructed from this database using a TPS-based spatial interpolation model for correspondence finding. Reconstruction is cast as a probabilistic estimation problem: given the surface data of a skull, find the most probable model instance such that the associated soft tissue thicknesses at the anatomical landmarks match the distances of the model surface to the skull surface. An EM-approach is followed to deal with outliers. This reconstruction method was tested on a database of 12 subjects for which both a 3D photo and a 3D CT was obtained.
A second reconstruction method is based on a database of volumetric CT images, which are segmented into hard- and soft- tissue volumes. For each of these volumes, Distance Transform (DT) images (implicit descriptions of the associated surfaces) are constructed. Facial reconstruction of a single skull is obtained by (B-spline) warping each skull DT image in the database to the target skull DT image, applying the warp to the corresponding soft tissue DT volumes and (attribute-weighted) averaging of the warped soft tissue DT volumes.