A complementary study for the evaluation of face recognition technology

Abstract

In 2010, National Institute of Standard and Technology (NIST) of the U.S. published “Report on the Evaluation of 2D Still-Image Face Recognition Algorithms (MBE 2010 Still Face).” The report mentions that there has been a remarkably huge improvement in the area of face recognition technology from the start of FERET (FacE REcognition Technology) program in 1993 up to 2010. While MBE 2010 Still Face is considered to be one of the best references in choosing appropriate face recognition algorithms from various kinds of software programs in the world, several points seem to be missing that need to be taken into consideration in the evaluation of recognition accuracy when face recognition technology is made use of in criminal investigations. They are the evaluation of the influence coming from (a) longer lapse of time (15-year aging difference), (b) shooting angles (vertical and horizontal), (c) change of face expression (smiling and laughing), and (d) accessories (cap, sunglass, mustache and so on). As the images taken by CCTVs on streets aren't always ideal mug shots, these points are also crucial in selecting the best face recognition algorithms as a tool to fight against crimes. Police Info-Communications Research Center (PICRC) attempts to evaluate the accuracy of face recognition technology by choosing some of the representative face recognition algorithms mentioned in MBE 2010 Still Face. PICRC has certain image database that stores two groups of full-faced photographs of people taken at intervals of 15 years. For instance as for the evaluation of the point (a), after the representative face recognition algorithms compared the photographs of the people with those of their former selves already stored in the database step by step, the degree of face recognition accuracy were verified. It is confirmed that the latest face recognition algorithms are hardly influenced by the four points ((a)-(d)) mentioned above. This result can conclude that the analyses made in MBE 2010 Still Face should be reliable enough even for police organizations to choose suitable face recognition algorithms for criminal investigations.