An Efficient Iris Recognition Technique using CNN and Vision Transformer

Abstract

The usage of biometric identification has increased in recent years, with numerous public and commercial organizations incorporating biometric technologies into their infrastructures. One of the technologies is iris recognition which has been used as a biometric recognition compared to other modalities to combat identity abuse due to its ability to eliminate risk of collisions or false matches even when comparing large populations. The use of CNN is proven to provide high accuracy; however, this technology involves the need for a large dataset and higher computational cost. Therefore, this study uses a combined model of Convolutional Neural Network (CNN) and Vision Transformer (ViT) in identifying and verifying an iris image. By using the proposed learning rate, it proves that the novel hybrid model is capable to achieve up to 93.66% accuracy in recognizing iris images. The cross-entropy loss function was implemented to reduce the loss and it was able to predict the class label more correctly. In addition, the model was thoroughly tested on three publicly available iris databases, achieving satisfactory iris recognition results. Furthermore, this model has the potential to be used in other biometrics such as face and retina recognitions.