Performance analysis of NASNet on unconstrained ear recognition

Abstract

Recent times are witnessing greater influence of Artificial Intelligence (AI) on identification of subjects based on biometrics. Traditional biometric recognition algorithms, which were constrained by their data acquisition methods, are now giving way to data collected in the unconstrained manner. Practically, the data can be exposed to factors like varying environmental conditions, image quality, pose, image clutter and background changes. Our research is focused on the biometric recognition, through identification of the subject from the ear. The images for the same are collected in an unconstrained manner. The advancements in deep neural network can be sighted as the main reason for such a quantum leap. The primary challenge of the present work is the selection of appropriate deep learning architecture for unconstrained ear recognition. Therefore the performance analysis of various pretrained networks such as VGGNet, Inception Net, ResNet, Mobile Net and NASNet is attempted here. The third challenge we addressed is to optimize the computational resources by reducing the number of learnable parameters while reducing the number of operations. Optimization of selected cells as in NASNet architecture is a paradigm shift in this regard.