Domain Adaptation for Unconstrained Ear Recognition with Convolutional Neural Networks

Abstract

Ear recognition has gained attention within the biometrics community recently. Ear images can be captured from a distance without contact, and the explicit cooperation of the subject is not needed. In addition, ears do not suffer extreme change over time and are not affected by facial expressions. All these characteristics are convenient when implementing surveillance and security applications. At the same time, applying any Deep Learning (DL) algorithm usually demands large amounts of samples to train networks. Thus, we introduce a large-scale database and explore fine-tuning pre-trained Convolutional Neural Networks (CNN) to adapt ear domain images taken under uncontrolled conditions. We built an ear dataset from the VGGFace dataset by profiting the face recognition field. Moreover, according to our experiments, adapting the VGGFace model to the ear domain leads to a better performance than using a model trained on general image recognition. The efficiency of the trained models has been tested on the UERC dataset achieving a significant improvement of around 9\% compared to approaches in the literature. Additionally, a score-level fusion technique was explored by combining the matching scores of two models. This fusion resulted in an improvement of around 4\% more. Open-set and close-set experiments have been performed and evaluated using Rank-1 and Rank-5 recognition rate metrics

Keywords

• Ear recognition
• CNN
• Mask-RCNN
• VGG16
• VGGFace
• transfer learning