Czajka, A; Bowyer, KW
This paper presents experimental results showing uneven distributions of selected iris image quality metrics in the consecutive attempts in a biometric system that allows for multiple attempts to complete a transaction. We consider three iris image quality metrics that can be influenced by user behavior: usable iris area, motion blur and margin adequacy. The quality metrics are used to judge the overall quality of the iris image and accept or reject it on each attempt. The experiment simulates a typical physical access scenario with a maximum of three attempts in one transaction. One conclusion is that subjects rejected on the first attempt do, on average, improve the quality of their iris image on their second attempt. If their second image is rejected, the average quality improvement on the third attempt is lower. A second conclusion is that the probability of a subject being rejected on the second try is higher in average than the probability calculated for all subjects delivering their first samples. The latter finding contrasts with a common belief that each try in a single transaction can be assumed to be a draw from the same authentic distribution (and hence the rejection probabilities are equal in each try). A third interesting and surprising observation is that improvement of sample quality is higher for women than for men. To our knowledge, this paper presents the first research explaining the nature of multi-attempt iris recognition system and delivers conclusions that suggest that the default understanding of this process is too simplistic.