Measurement of the information for identification in iris images

Performance of iris recognition systems strongly depends on uniqueness of irises. One possible way to estimate the uniqueness is to measure the amount of identifying information contained in irises. No proper identification can be done if the quality of involved samples is poor. In our study, we verified human ability to assess iris sample quality and concluded that their intuitive quality assessment may not be sufficient for biometric identification applications. After image quality is assured, the biometric information measurement may be addressed. We studied the approach proposed by Daugman that measures the amount of information in irises. Experiments and statistical analysis showed that it does not provide a useful way to measure an amount of identifying information. An alternative algorithm based on relative entropy was developed to measure information important for identification. We used it to detect the most informative regions of the iris and to understand dependencies within it