טכניון מכון טכנולוגי לישראל
הטכניון מכון טכנולוגי לישראל - בית הספר ללימודי מוסמכים  
M.Sc Thesis
M.Sc StudentMichael Bronstein
SubjectThree-Dimensional Face Recognition
DepartmentDepartment of Computer Science
Supervisor Full Professor Kimmel Ron


Abstract

The term biometrics refers to a variety of methods that attempt to uniquely identify a person according to a set of such features. It is well-known that some characteristics or behavior patterns of the human body are strictly individual - a few such examples include the DNA code, fingerprints, structure of retinal veins and iris, individual's written signature or face. One of the oldest written testimonies of a biometric technology and the first identity theft dates back to biblical times, when Jacob fraudulently used the identity of his twin brother Esau to benefit from their father's blessing. In this dissertation, we focus on face recognition, which is probably the most natural biometric identification method used by the humans from the dawn of the civilization. One of the greatest challenges, the “Holy Grail” of automatic face recognition is dealing with facial expressions. The ability to recognize human faces in the presence of expressions is necessary, for example, in commercial face recognition systems. We introduce a geometric framework, treating faces as deformable Riemannian surfaces. Facial expressions are modeled as isometries of the facial surface, and the recognition problem is reduced to finding similarity between isometric surfaces. Next, we represent the Riemannian surface as a subset of some convenient low-dimensional space such that the original intrinsic geometric structure is preserved. This procedure, known as isometric embedding, allows to create an expression-invariant representation of the face. We discuss several examples of embedding spaces with flat and spherical geometry, and describe numerical optimization algorithms to perform the embedding. Finally, we demonstrate an implementation of our approach in an accurate three-dimensional face recognition system that is capable of distinguishing between identical twins.