We are developing solutions that will enhance the performance of face recognition engines by reducing accuracy errors associated with longitudinal aging. We have published articles on mitigating aging effects on commercial FR engines. (See pub lists.) We collaborate with FR developers and manufacturers to assist them in designing a more robust system to aging. Our long experience in this area has imparted lots of wisdom regarding feature selections and matching enhancements. We have extended the work for adult aging to that of children with support by researchers at Concordia University.
Most FR work has been evaluated on adult faces, and adult faces close in time (few days to few years between gallery and probe). We will soon begin a large collection effort focused on children and teenagers (pre-adults) to address the concerns of FR efficacy on this group. We hope to make this data collection available to the research community in the future.
Synthetic Face Aging
Synthetic aging is our bread-and-butter, we have developed several algorithms for synthetically aging and/or de-aging a person. Our techniques are grounded in the science of craniofacial morphology. We have notable experts on our team or in collaboration that we leverage to increase our understanding of the mechanisms associated with aging. Our techniques are built upon this understanding of the fundamental principles of aging for adults and youth. Therefore, we are able to create realistic synthetic images of future / past faces based on the individual drivers of a person, this is known as idiosyncratic aging. Most techniques employed today are based upon general aging trends, which applies these aging trends to all persons equally. This approach cannot account for the differences in aging due genetics, environment and/or behavior.
We are investigating the use of low cost iris sensors and their efficacy as a stable verification platform that can be integrated into small appliances. We are also exploring, in concert with our CASIS partners, the ability to perform long range iris, bit-code reduction and mitigation of fragile bits, and fusion of iris with skin region around the eye (periocular). We are developing models of aging around the periocular and have partnered with a respected plastic surgeon to understand the dramatic changes that occur in this region due to aging and reconstructive surgery.
We are actively looking for sponsors to explore our ideas in craniofacial micro / macro gestures. We have some very promising results in this area for identity verification and trustworthiness.
We have developed several algorithms for age estimation that are cutting edge. These algorithms continue to push the performance on standard databases like FG-NET. We have published results with mean absolute error (MAE) on FG-NET at 4.37 years (with Concordia University). Our latest results on FG-NET resulted in rates below 4.0 years in cross-validation testing. We have also developed algorithms that show promise against ethnically diverse data sets.
Mobile biometrics are defined as the use of biometric techniques on mobile devices. Mobile devices are the perfect platform for many biometric systems as they have very powerful processors, large storage, high resolution cameras, and a small, portable footprint. Mobile biometrics encompass the development of mobile devices to acquire biometric signals, software algorithms for identification and verification, and data stores to house biometric data for comparison. Our current research focus is on the integration of mobile devices with face recognition. We have developed some prototype systems over the last few years to incorporate face recognition on camera phones (circa 2007) and the integration of texting with face recognition.
We have leveraged feature sets from age estimation for gender classification. We are building unified set of features that can be used for all demographic information extraction. Gender classification teamed with age estimation is a powerful tool for marketing and access control.
Race classification is an extremely difficult topic due to the cross-pollination of racial groups (known as admixture). However, we are work on developing coarse race classification systems.