Residual Network Deep Learning Model with Data Augmentation Effects in the Implementation of Iris Recognition

Abstract

Research in iris recognition using deep learning methods has gained significant traction in recent times, emerging as a robust biometric identification approach. Iris patterns encompass diverse elements like color, vascular structure, and texture, which combine uniquely and pose challenges for forgery or replication. ResNet, a specialized Deep Learning model designed for object recognition, has demonstrated remarkable performance when trained on the ImageNet Dataset. However, iris images pose a considerable challenge for recognition due to their intricate features, which are hard to discern through conventional observation or content-based feature extraction methods. This study seeks to assess the efficacy of ResNet-34 and ResNet-50 in iris recognition. Additionally, a data augmentation process will be implemented to expedite the training process and enhance accuracy. it is evident that ResNet-34 achieves a higher accuracy than ResNet-50. Specifically, ResNet-34 yields a test accuracy of 0.751 for the original dataset and 0.768 for the augmented dataset, whereas ResNet-50 achieves an accuracy of 0.73 for the original dataset and 0.747 for the augmented dataset.