Experimental Evaluation of Iris Recognition

Iris is an important biometric method with high reported accuracy. However, current iris recognition systems require substantial user cooperation in the image acquisition. Relatively little is known about how iris recognition might perform with less stringent control of image quality. We have re-implemented a Daugman-like iris matchingmethod, and evaluated its performance on an image dataset of over 12,000 images from over 300 persons, with iris images of different qualities. We find an overall rank-one recognition rate using of 89.64%. Poor quality images account for most of the instances of incorrect recognition. Inaccurate segmentation is also a key problem. These results show that greater flexibility in use of iris recognition will require further work on handling images of non-ideal quality. We also explore the use of multiple images for representing a person.