Face reconstruction from skull based on Least Squares Canonical Dependency Analysis

Abstract

Face reconstruction from skull, called as Craniofacial Reconstruction (CFR), is a useful technique to identify an unknown decomposed corpse if no other evidence is available. Traditional manual methods greatly depend on the experience of sculptors, so that the results are highly subjective, and the whole process is time consuming. Recent years, 3D data acquiring technology becomes consummate, and machine learning techniques raise a tidal wave in academia and industry. Researchers turn to finding computer aided solutions, especially the supervised machine learning technique, for craniofacial reconstruction. Least Squares Canonical Dependency Analysis (LSCDA) is a dimension reduction method, which aims at finding subspaces where the dependency measured by Least Squares Mutual Information (LSMI) of two variables reaches maximum. This paper proposes a new method for craniofacial reconstruction based on LSCDA. First, two statistical shape models for skull and skin are constructed respectively by Principle Component Analysis (PCA). Then the subspaces of maximum dependency of face and skull are extracted in the shape parameter spaces via LSCDA. Finally, according to such dependency, the relationship model between skulls and skins is established by Least Squares Support Vector Regression (LSSVR), which is used to reconstruct the facial appearances for an unknown skull. Experiment results show that the proposed method is effective.