Deep Learning Frameworks for Off-Angle Iris Recognition

Abstract

Although iris is known as one of the most accurate, distinctive, and reliable biometric identification, the accuracy of iris recognition depends on the image quality and it is negatively affected by several factors such as gaze angle, occlusion, and dilation. Since standoff iris recognition systems are much less constrained than traditional systems, the iris images captured are likely to be non-ideal, off-angle, and dilated. In this paper, we present a deep learning algorithm to improve performance of off-angle iris recognition in traditional and untraditional iris recognition frameworks. As a main contribution, this approach will allow us to study the effect of the gaze angle in ocular/periocular biometrics and fuse the information in different off-angle iris images. Using convolutional neural networks (CNNs), we first investigate traditional iris recognition framework with segmentation, normalization, and CNN based encoding and matching. In nontraditional framework, we use the iris images without segmentation and normalization to investigate the effect of periocular regions. We train and test our deep learning based approach with frontal and off-angle iris images in different sizes and types such as original, cropped, segmented, and masked images. Based on results from our real off-angle iris image dataset with 52 subjects, the proposed method improved the recognition performance compared with traditional off-angle iris recognition approaches.