Edge Artificial Intelligence

6Comparative Analysis of Pruning Conventional Machine Learning or Deep Learning Frameworks Utilizing Discrete Wavelet Transform for Iris Biometrics

Divyang Jadav*, Aditee Moudgil, Nilesh Choudhary and Shruti Daw

SVKM’s NMIMS, STME, Navi-Mumbai, India

Abstract

Today, in the post-COVID-19 environment, contactless acquisition of biometric data has become popular for hygiene purposes and better accuracy. Compared to other biometric traits, the iris is considered the most secure and inherently touch-free modality, making it suitable for this research’s focus. It employs state-of-the-art machine learning and deep learning models to detect features of individuals for verification and identification tasks. Conventional deep learning networks, although highly accurate, come with issues like dependency on large datasets, dense frameworks, computational incapability, inter-data set variability, etc. Furthermore, in most of the analyzed methods, the hyperparameter tuning is still somewhat ad hoc and does not rely on strong mathematical justification. These problems make them less desirable for resource-scarce edge nodes, thus requiring model pruning without loss of accuracy. This study presents a hybrid framework for iris recognition that uses the discrete wavelet transform (DWT) employing the Haar wavelet in the preprocessing step. Critical features extracted using DWT are used to train support vector machine (SVM), convolutional neural networks (CNN) and ResNet50 models. Performance ...