IREX III Supplement I: Failure Analysis. NIST Interagency Report 7853.

摘要

Iris recognition has the potential to be extremely accurate, but like many other biometrics, it is highly dependent on the quality of the sample data. Occlusion of the iris, off-axis gazes, blurred images, and iris rotations are common problems that can make recognizing individuals more difficult. A fraction of the Iris Exchange (IREX) III dataset consisted of poor quality images that reduced overall performance in the evaluation. For single eye searches, the matching algorithms were never able to achieve an identification rate better than 98.3% at any reasonable decision threshold. It is fair to assume that some large fraction of those missed searches are the result of poor quality on the part of either the search image or its corresponding mate in the enrollment set. The poor quality images often stem from systemic problems with the capture and data handling process. The purpose of this failure analysis is to identify common sources of poor sample quality and to provide best practice recommendations for how to improve the quality of captured samples. The analysis began with the construction of a special failure set of images, which contained those images that were a common source of error for several matching algorithms. Strictly speaking, the failure set contained a set of image pairs, where each pair consisted of a search image and a corresponding mate that was used for enrollment. A thorough manual inspection was then performed on each image in order to identify specific cause(s) of failure. Section 2 describes the construction of the failure set and manual inspection process. Section 3 explores each of the common sources of poor sample quality and provides recommendations on how to adjust the capture protocol or data management practices to reduce the likelihood of their occurrence. Appendix A summarizes the best practice recommendations provided in the previous subsections.