An Accurate Iris Recognition System Based on Deep Transfer Learning and Data Augmentation

Iris recognition is a highly secure biometric identification method that utilizes the distinct and intricate patterns of the iris, which remain consistent throughout a person's lifetime. Remarkably, it can even distinguish between identical twins, highlighting its precision and reliability. However, despite these advantages, challenges such as reflections, occlusions, varying lighting conditions, blur, and other environmental factors can significantly degrade the performance of iris recognition systems. Deep Convolutional Neural Networks (CNNs), known for their success in image recognition, often require numerous parameters, increasing computational time and resource demands. To address these challenges, we propose an advanced iris recognition system based on Deep CNNs. Our system utilizes YOLOv4-tiny for accurate iris region detection, and for the recognition task, we introduce a deep CNNs model inspired by the pre-trained Xception, which contains fewer than three million parameters. To effectively train this deep CNN on small iris datasets, we apply Data Augmentation and Transfer Learning techniques, which improve accuracy while reducing the risk of overfitting. Through two-fold cross-validation, our system achieves impressive accuracy rates of 99.81 % and 99.64 % on the CASIA-Iris-Interval and IITD datasets, respectively. Our method demonstrates a clear advantage over existing methods in terms of accuracy and reliability, making it a promising solution for various applications such as border control, law enforcement, and access control.