2D/3D Face Recognition

摘要

Traditional biometric research over the past 20 years has focused primarily on the ability to recognize subjects in two scenarios: identification and verification. Verification is the process that attempts to confirm the identity of a given individual (i.e., John Smith presents his passport to the gate-agent to verify that he really is John Smith). This process is usually very quick and easy to perform on the computing power available today. Conversely, identification is a process that attempts to locate an unknown subject given a list of individuals (i.e. US Terrorism Watch List). This process is very time consuming because each member on the watch list must be compared to the subject in question. In both of these processes, we have generally assumed that the images will display cooperative subjects in well- defined lighting environments. In the real world, this is not the case. While it is assumed that subjects will be cooperative during an initial enrollment phase (such as getting a driver's license, going through a checkpoint, or walking through an x-ray machine), this is usually not the case for future acquisition attempts. When attempting to identify an individual at a later time (especially at a distance and without their knowledge in a covert scenario), these 'probe' images are generally nonfrontal facing, in uncontrolled lighting, and at an unknown distance. When standard recognition algorithms are executed on this variable dataset, results are typically very poor. In this project we developed a complete solution for 3D based 2D facial recognition that leverages existing research and previously developed code by domain experts at the University of Southern California and Progeny Systems. To verify performance, we employed multiple public facial data sets acquired for the purpose of 2D and 3D facial recognition - most recently MultiPIE - which varied pose, illumination, and expression.