Abstract:
Ear biometrics has risen as a promising methodology for human identification,
due to the ear’s uniqueness, permanence, and constrained variability over time.
This thesis centers on enhancing ear recognition systems utilizing deep learning
techniques, particularly by leveraging convolutional neural networks. The proposed
approach includes implementing and fine-tuning the ResNet50 architecture utilizing
the AMI Ear dataset, which contains ear images captured under diverse conditions
and postures. Data preprocessing and augmentation methods are applied
to enhance model robustness, and training is conducted with careful tuning
to adapt to the specific characteristics of ear pictures. The experimental
results demonstrate that the proposed ResNet50-based model achieves improved
recognition performance, reaching an accuracy of 99.29%, outperforming previous
methods. This work demonstrates the potential for deep convolutional models in
ear biometrics and sets the stage for future applications of multi-view fusion, large
datasets, and combined modalities with other biometric traits.
