Avaliação do impacto da qualidade da imagem no processo de reconhecimento biométrico da íris baseado na norma ISO/IEC 29794-6:2015

Abstract

The use of biometrics to identify and authenticate people is already present in several areas, such as the banking system, electronic point registration and even on mobile devices. The basic principle of biometrics is to identify a person by means of a unique characteristic that distinguishes them from other individuals and can be done through digital fingerprinting, facial recognition, voice recognition, among other forms. Another technique, considered one of the most reliable and in great expansion, is the biometry through the texture of the iris, but its accuracy depends directly on the quality of the images of the iris. In this context, this study aims to evaluate the impact of iris image quality on the performance of a biometric system. For this purpose, the recently published ISO / IEC 29794-6: 2015 was adopted, with the proposal to define relevant quantitative terms and methodologies to characterize the quality of the iris images and to evaluate their potential for high biometric reliability in decision making in the recognition process. The study was applied to three different iris databases (CASIA, IITD and UFLIRIS) using a free iris recognition software, Osiris v4.1, which was widely referenced in other studies. The UFLIRIS database, developed as part of this research, had the participation of 26 individuals, 1728 images were obtained in different positions and quality levels, assisting the research in the results obtained and that will be made available in conjunction with their segmented binary masks manually. The experimental results in the UFLIRIS database show that the Q1 metric, which refers to the visible area of the iris texture, is the metric that has the greatest impact on the performance of the system, considering the F-Meausre, obtaining a improvement from 0.5026 to 0.84373. On the other hand, when considering the Q9 metric, which evaluates the adequacy of the image margin, it would be necessary to remove more than 91.42% of the images in the CASIA database and 99.36% in the IITD database that were below the reference value recommended by the standard. It was verified that even without the removal of these images, it was possible to obtain performances of 0.98876 in the CASIA base and 0.96638 and in the IITD base, demonstrating that adopting this metric may result in the unnecessary removal of many images. Finally, the rotation between comparative images has proved to be a great problem, often neglected by the iris recognition algorithms and directly affecting the results obtained, mainly with the increase in False Non-Match Rate (FNMR) values, as for example, in the UFLIRIS data set, which included images with a high rotation difference, 54% of FNMR was verified and with the removal of these images the value of FNMR was reduced to 13%.