COSC 460 Improving Face

Face recognition has long been an area of great interest within computer science, and as face recognition implementations become more sophisticated, the scope of real-world applications has widened. The field of genealogy has embraced the move towards digitisation, with increasingly large quantities of historical photographs being digitised in an effort to both preserve and share them with a wider audience. Genealogy software is prevalent, but while many programs support photograph management, only one uses face recognition to assist in the identification and tagging of individuals. Genealogy is in the unique position of possessing a rich source of context in the form of a family tree, that a face recognition engine can draw information from. We aim to improve the accuracy of face recognition results within a family photograph album through the use of a filter that uses available information from a given family tree. We also use measures of co-occurrence, recurrence and relative physical distance of individuals within photos to accurately predict their identities. This proposed use of genealogical and contextual data has shown a 26% improvement in accuracy over the most advanced face recognition technology currently available when identifying 348 faces against a database of 523 faces. These faces are extracted from a challenging dataset of 173 family photographs, dating back as far as 1908. Acknowledgments I’d like to thank Glen Cameron from NEC NZ for his assistance in setting up and using NeoFace, Janice Newcombe for her help gathering the photographs and genealogical data needed for this research, Christoph Bartneck from the HIT Lab NZ, who initiated this research, and Richard, for being a great supervisor.