OcularAge: A Comparative Study of Iris and Periocular Images for Pediatric Age Estimation

TL;DR

This work investigates pediatric age estimation from ocular biometrics, comparing iris and periocular inputs using a longitudinal NIR dataset of 288 children aged 4 to 16 collected over eight years. A multi-task CNN framework jointly predicts age group and exact age across six architectures, evaluated on two sensor devices and deployed on high-performance and VR hardware. The study finds periocular (eye) images consistently outperform iris images, achieving a mean absolute error of 1.33 years and age-group accuracy of 83.82%, with strong cross-sensor generalization for select models and real-time on-device inference on VR headsets. These results establish a longitudinal benchmark for pediatric ocular age estimation, demonstrating the feasibility of privacy-preserving, on-device age checks in child-centric VR applications and informing robust, sensor-robust biometric system design.

Abstract

Estimating a child's age from ocular biometric images is challenging due to subtle physiological changes and the limited availability of longitudinal datasets. Although most biometric age estimation studies have focused on facial features and adult subjects, pediatric-specific analysis, particularly of the iris and periocular regions, remains relatively unexplored. This study presents a comparative evaluation of iris and periocular images for estimating the ages of children aged between 4 and 16 years. We utilized a longitudinal dataset comprising more than 21,000 near-infrared (NIR) images, collected from 288 pediatric subjects over eight years using two different imaging sensors. A multi-task deep learning framework was employed to jointly perform age prediction and age-group classification, enabling a systematic exploration of how different convolutional neural network (CNN) architectures, particularly those adapted for non-square ocular inputs, capture the complex variability inherent in pediatric eye images. The results show that periocular models consistently outperform iris-based models, achieving a mean absolute error (MAE) of 1.33 years and an age-group classification accuracy of 83.82%. These results mark the first demonstration that reliable age estimation is feasible from children's ocular images, enabling privacy-preserving age checks in child-centric applications. This work establishes the first longitudinal benchmark for pediatric ocular age estimation, providing a foundation for designing robust, child-focused biometric systems. The developed models proved resilient across different imaging sensors, confirming their potential for real-world deployment. They also achieved inference speeds of less than 10 milliseconds per image on resource-constrained VR headsets, demonstrating their suitability for real-time applications.

Figures (9)

• Figure 1: Example NIR eye images captured using two different iris imaging devices. The left column shows images collected with the IG-AD100, while the right column shows images from the iCAM T10. Rows correspond to ages 9, 10, 11, and 12 (top to bottom), illustrating consistent cross-sensor capture and longitudinal variation.
• Figure 2: Distribution of images by subject age.
• Figure 3: Example of the iris pre-processing pipeline. From top to bottom: the original eye image, the corresponding normalized iris image generated using OSIRIS, and the binary mask indicating the valid iris region. The mask is used as a second channel during model training to guide the network’s attention to meaningful iris features.
• Figure 4: Overview of the model architecture. Eye-based models take a 1-channel grayscale input, while iris-based models use a 2-channel input comprising the normalized iris image and its binary mask. All models share a modified input stem adapted from pretrained ImageNet backbones, followed by a shared feature extractor (backbone) and two output heads: one for age group classification and one for exact age regression.
• Figure 5: Grad-CAM visualizations from the eye-based MobileNetV3 model. Each image is annotated with the ground truth class (GT), predicted class (Pred), and classification confidence (Conf). The left column shows correctly classified examples with focused attention on periocular regions such as eyelid folds, scleral boundaries, and lash contours features known to evolve with age. The right column shows misclassified samples, where the model attention is scattered or misaligned, often falling on noninformative regions such as background skin or upper brow. These failure cases highlight the model’s sensitivity to image ambiguity and its reliance on high-quality anatomical cues for accurate prediction.