Rekonstruktion von Gesichtern mittels Freiform-Deformation

Master-Thesis

The main goal of this investigation is to design and develop an efficient strategy for reconstructing the mean face features from raw skull Dicom images. A reconstruction of a face basically associates a determined muscle and face architecture to a particular skull structure and applications go from forensic reconstructions up to plastic surgery plannings. The problem can be split into two components: Development of computer graphic tools and a non linear predictor model. Computer graphic techniques are herein used for extracting the raw data from the Dicom image, reconstruction, which includes interpolation for data representation, and display. The computer graphic development goes necessarily by surface reconstruction from the raw data and implementation of efficient techniques for data handling. Finally the graphic problem is complemented by strategies able to deform large mesh surfaces and linked to general restrictions such as specific curvature conditions or particular data relationships. Free Form Deformation has the advantage that it can both handle smoothness issues and adapt to specific face features with higher curvature. For this approach we load a skin surface former reconstructed by Dicom images and apply Free Form Deformation to that surface to adapt it properly to the investigated skull. It should be pointed out that any deformation strategy must be leaded by a model which can take into account the anatomical issues. Such a model will be based on heuristical or anatomical knowledge but will use current machine learning techniques, which are more adaptable to non-linearities. Among the techniques we count to assess we have the SVM (Supported Vector Machine) against a naive Bayes, which is simply the simpler Baye's theorem implementation. This is very convenient in problems in which one has no metric defined on the parameter space and will be compared with an approach (SVM) which has a strong implicit concept of metric. Because of portability reasons we use Java for the implementation. For the graphic part OpenGL is used which can be included in Java by the Jogl library.

Advisor(s)