Few-shot learning for ear recognition

Ear recognition is a popular field of research within the biometric community. It plays an important part in automatic recognition systems. The ability to capture image of the ear from a distance and perform identity recognition makes ear recognition technology an attractive choice for security application as well as other related applications. However, datasets of ear images are still limited in size, while in other biometric modal communities, like face recognition, they possess large datasets and the most of them are collected in uncontrolled condition. As a result, deep learning technology still cannot yield satisfactory result in ear recognition area. In this paper, we tackle ear recognition problem by using few-shot learning based methods. We explore different methods towards model training with limited amounts of training data and show that by using them, with the help of data augmentation, the model can be flexible and can quickly adapt to new identity to perform fast recognition. The result of our work is the first few-shot learning based work to ear recognition. With our work we are able to significantly improve the accuracy of 23% on a regular dataset, and even 21% on a challenging dataset that is collected from the web, which is comparable with state-of-the-art in ear recognition area.