Iris recognition system (irs) using deep learning technique
Loading...
Date
2019-06
Authors
Yow, Sue Chin
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Biometric authentication becomes sophisticated with the advance of technology, especially progressive image processing, and computational capabilities. Iris recognition through human iris organ is one of the popular biometrics as it is promising higher accurate return in machine vision, reliability and simpler as compared to other traits. Machine Learning depends on the input fit in, do classification on the sample data. Finally, predict the output based on the probability. For the previous classical recognition system technique, accurate iris segmentation is the crucial part to guarantee high accuracy for Iris Recognition System (IRS). Hence, the requirement of the image dataset has to acquire under specific conditions, else it might be failed in iris segmentation as well as the hyperparameters set or algorithms applied unsuitable for it. Manual hyperparameter tuning on Machine Learning model may take time and failure if not fully understand the algorithms and feature of datasets work with. In this thesis, the Transfer Learning method is proposed to capitalize pretrained Convolutional Neural Network (ConvNet) model introduced in the ImageNet Large Scale Visual Recognition Competition (ILSVRC) on the IRS. Systematic analysis has been conducted to design an optimal deep network architecture to achieve high efficiency in feature extraction. AlexNet and DenseNet201 pre-trained model that poses different ConvNet architecture and layer depth were chosen and trained Support Vector Machine (SVM) for testing model transferability. CASIA-Iris-Interval V1 dataset is then re-trained on AlexNet and DenseNet201 model one by one. Finally, evaluation of the IRS performances after applying Data Augmentation and Bayesian Optimization. All the results recorded along the algorithm development process showed the success of proposed methodologies in gaining a higher performance algorithm. Undergo proposed methodology flow, AlexNet achieved an overall accuracy of 97.22% meanwhile DenseNet201 achieved an overall accuracy of 98.81%. Transferability of a pre-trained model on new target task is improved and meanwhile, the high recognition rate of the algorithm on small-size CASIA-Iris-Interval V1 iris image dataset is achieved.