On distinctiveness in ear biometrics

Abstract

Ear biometrics have developed rapidly in last decade. Ears have distinct advantages over face and fingerprint, such as invariant structure over time, and ear images can be captured without subject’s participation. There are also some considerations when ears are used as a biometric, such as rotation variance, varying illumination and occlusion by hair. Wider application has been hindered by these problems. Previous works show that human ears can be used for identification, gender classification, and age classification. In this thesis, we propose a new model-based approach to ear biometrics, which contains geometric features based on ear anatomy. The keypoints of our model are determined by scale-invariant feature transform (SIFT), and we consider the rotation of ear images under an affine transformation, by modelling the ear as a flat plane attached to the head. Then, we extend our model with image pre-processing step that using the force field transform to remove the noise.
We apply the model and fine-tuned convolutional neural networks on ear recognition, gender classification and ear symmetry. In ear symmetry, we address the question as to whether it is possible that given an image of one ear, a person can then be recognized from his/her other ear. Such a symmetry-based strategy could reduce constraints on applications of ear biometrics. To investigate symmetry, we compare one ear with a mirrored version of the other ear.
In addition, we consider the important parts of ear recognition, gender classification and ear bilateral symmetry on ear images, in these three cases we aim to determine the ear parts from which recognition is derived. For analysing the model-based, we use accuracies of different ear regions to evaluate the significant parts for ear recognition, gender classification and ear symmetry. Moreover, we are the first to apply the heatmaps on ear images to determine the contributions of different parts of ear, and this is the first study to analyse the differences between male and female. Also, we have compared the model-based method with deep learning, and the contributions of different parts based on different approaches.
Furthermore, we are the first to exploit ear for kinship verification, and we collect SOTEAR dataset for the kinship verification experiments. We compare the influence of father with that of mother by the accuracies of kinship verification.

More information

Identifiers

ORCID for Mark Nixon: orcid.org/0000-0002-9174-5934

Catalogue record

Export record