Unconstrained and Constrained Ear Recognition Using Deep Learning Architectures

For a long time now, a means to identify individuals by their physical traits has become crucial. Indeed, identification can be the difference between freedom and imprisonment, it can be the difference between a safer world or a more dangerous one. While each biometric has its own advantages and disadvantages, the ear, in particular, is one of, if not the most useful for forensic investigators because of its slowly changing nature, empowering them can mean empowering justice. This research proposes a new deep-learning model, that is trained and tested on both constrained (AMI) and unconstrained (AWE, EarVN) ear databases of individuals. The model takes an ear image as an input, then it outputs the predicted identity of the individual from the database it was trained on. The proposed approach takes advantage of various factors that affect a deep learning model’s accuracy. The most significant of which are the feature extractor, image augmentation, regularization, optimizer and fine-tuning. This research develops the model with these factors in mind and explores finding the best combination of these techniques to maximize the accuracy and robustness of the system. The developed approach has shown very promising results, consistently identifying at least 94.7% of individuals in the EarVN database and even higher in the other two. Keywords: Deep Learning, Computer Vision, Biometrics, Constrained Ear Recognition, Unconstrained Ear Recognition