Studying the Generalization Capability of the ResNet-50 model in Independent Iris Classification Tasks

Abstract

This study evaluate  a deep learning and transfer learning framework based on the ResNet-50 architecture to support iris biometric systems through multiple classification tasks.. In this paper, we propose training several distinct and independent models for three non-sequential classification tasks: image authenticity verification (fake/real), gender classification (male/female), and iris orientation determination (right/left). This design aims to evaluate the effectiveness of employing a unified backbone architecture for multiple classification tasks, enabling efficient handling of various classification challenges while maintaining system flexibility. Furthermore, this approach establishes a foundation for the future integration of individual models into a unified framework capable of simultaneously classifying multiple attributes. The experimental results demonstrated that the proposed models achieved promising classification performance across different tasks. The Real/Fake classification model obtained the highest accuracy, achieving approximately 99–100% validation accuracy, indicating strong generalization capability. In contrast, the Gender and Iris Orientation classification models showed relatively lower performance due to the higher variability in iris image characteristics, achieving validation accuracies of approximately 86% and 80%, respectively. Demonstrating the proposed framework's flexibility and strong generalization capability, thus increasing its potential for adoption and use in security and intelligent biometric systems.