U-Net convolutional neural network for multisource heterogeneous iris segmentation

Abstract

Accurate iris segmentation is a critical step in various applications, from biometric identification systems to ophthalmic disease diagnosis. Despite the large number of works that address this problem, iris segmentation of heterogeneous iris images acquired in different conditions is still a huge challenge. This work employed a modified U-net convolutional neural network architecture to segment iris region from heterogeneous eye images. The network was trained using the TEyeD dataset, the world’s largest heterogeneous publicly available dataset of eye images.The proposed method utilizes the U-Net architecture, known for its effectiveness in handling complex image segmentation tasks. The architecture is modified to accomplish the specific task. The experimental results show that the proposed approach achieves an IOU score of 95%, demonstrating promising results in terms of segmentation accuracy and computational efficiency. This performance is competitive or even better than the existing state-of-the-art techniques in iris segmentation, considering that in most cases the dataset used to train the network is not heterogeneous as the TEyeD dataset.Indeed, the study highlights the potential of deep learning techniques in improving the accuracy of iris segmentation, and the TEyeD dataset, which is a heterogeneous dataset in terms of acquisition devices employed and image quality, provides an excellent opportunity for researchers to further explore this topic. The findings of this research could have significant implications for various fields, including biometric identification systems, driver safety, and ophthalmology.