IREX III: Performance of Iris Identification Algorithms. NIST Interagency Report 7836

摘要

The Iris Exchange IREX III evaluation was conducted to measure the accuracy and speed of iris identification algorithms. The test was designed to have operational relevance. First, by using as many as 6.1 million images of 4.3 million eyes, the results are relevant to systems used for the full range one-to-many applications including de-duplication, benefits fraud, and token-less access and border control. Second the test ran algorithms on commodity PC-class blade computers running the LINUX operating system which are typical in central-server applications. Third, the test applied algorithms to archival operational data, thereby mimicing a real-world task. The algorithms were invoked to do one-to-many searches in a database of enrolled iris images to produce lists of candidate identities sorted in increasing order of dissimilarity value. Two kinds of searches are executed. The first, searches with an enrolled mate, allow measurement and reporting of the core false negative identification rate, i.e. the miss rate. The second, searches for which there is no enrolled mate, support measurement of false positive identification rate, more commonly known as false alarm rate. These quantities are estimated as a function of enrolled population sizes. While the test is primarily useful for comparative evaluation of algorithms, it is also pertinent to algorithm developers and those concerned with operational deployment of iris identification.