A Unified Low-Latency YOLOv5–YOLOv8 Framework for Real-Time Iris–Pupil Detection in Biometric and Clinical Systems

Abstract

Real-time detection of the iris and pupil is crucial for applications such as biometric authentication, medical diagnosis, and eye tracking. This paper presents an efficient approach using YOLOv5s, YOLOv8s, and YOLOv11s for iris and pupil detection and subsequent iris-to-pupil ratio estimation. The proposed method exploits the speed and accuracy of YOLO architectures to achieve robust segmentation and reliable ratio computation under varying illumination and occlusion conditions. The models are trained on a diverse dataset with data augmentation to improve generalization across different eye shapes and lighting scenarios. The training pipeline incorporates image preprocessing, transfer learning with pre-trained weights, and evaluation using standard performance metrics. Experimental results demonstrate testing accuracies (mAP@50) of 99.5%, 99.7%, and 99.5% for YOLOv5s, YOLOv8s, and YOLOv11s, respectively, with YOLOv8s achieving the best overall performance on the proposed dataset. Compared with related methods, the proposed approach offers superior speed and reliability for real-time applications in security, medical diagnostics, and gaze tracking. Future work will focus on optimizing model architectures, expanding dataset diversity, and enhancing adaptability to varied eye structures, further advancing real-time eye-tracking and physiological monitoring systems.