Hybrid CNN-ViT integration into Siamese networks for robust iris biometric verification

Abstract

Iris recognition has emerged as a critical biometric verification method, valued for its high accuracy and resistance to forgery. However, traditional convolutional neural network (CNN)-based models, despite their strength in extracting local iris features, struggle to capture global dependencies, which limits their generalization across different datasets. Additionally, conventional classification-based approaches struggle to accurately verify new individuals with limited training data. Thus, this study proposed a hybrid CNN-vision transformer (CNN-ViT) model within a Siamese network to enhance one-shot learning capability by combining CNN’s local feature extraction with vision transformers (ViT’s) global attention. To evaluate its performance, the hybrid model was compared with VGG16 and ResNet under the same training conditions for 20 epochs. VGG16 and ResNet rely on pre-trained models, whereas the hybrid CNN-ViT model is specifically designed to achieve this task with an increment to 98.9% training accuracy, surpassing the TinySiamese model's benchmark accuracy. It also attained a recall of 75%, demonstrating strong sensitivity in correctly identifying positive matches. The hybrid model maintained an excellent balance between learning and generalization by employing the binary cross entropy (BCE) loss function. These findings contribute to the development of efficient iris recognition systems, paving the way for advanced biometric applications in financial transactions, border control and mobile security.