New methods for 3D iris scanning for multiple 2D visible-light images

Abstract

Iris recognition is one of the most successful biometric methods; however, it uses 2D images for the analysis when the iris is in fact a 3D muscular structure. Those muscular fibers create a relief in the iris surface iris, which is what we propose to characterize in a 3D model. The additional depth information aims to increase iris recognition performance and has potential applications in ophthalmology. In this Doctoral thesis, we developed and compared the performances of two different approaches to 3D iris scanning using separately Structure from Motion (SfM) and Convolutional Neural Networks (CNN). To train the proposed CNN architecture, irisDepth, we captured 26,520 images from 120 subjects. The SfM method produced 11,105 3D points in average while the CNN method produced 6 more at 65,536. The resolution of SfM and CNN were 11µm and 17.7µm respectively. The average error between a ground-truth Optic Coherence Tomography and the corresponding slice in the 3D model was 123µm for SfM and 77µm for CNN. Thus, the CNN method increased accuracy in 60% with respect to SfM. Finally, the 3D models increased iris recognition performance 68% with respect to the standard iris code in a dataset of 50 subjects and 2,000 images.