Exploring Complementary Features for Iris Recognition on Mobile Devices

Iris recognition on mobile devices is challenging due to a large number of low-quality iris images acquired in complex imaging conditions. Illumination variations, low resolution and serious noises reduce the distinctiveness of iris texture. This paper explores complementary features to improve the accuracy of iris recognition on mobile devices. Firstly, optimized ordinal measures (OMs) features are extracted to encode local iris texture. Afterwards, pairwise features are automatically learned to measure the correlation between two irises using the convolutional neural network (CNN). Finally, the selected OMs features and the learned pairwise features are fused at the score level. Experiments are performed on a newly constructed mobile iris database which contains 6000 images of 200 Asian subjects. Their iris images of left and right eyes are obtained simultaneously at varying standoff distances. Experimental results demonstrate OMs features and pairwise features are highly complementary and effective for iris recognition on mobile devices.