Domain Adaptation for Unconstrained Ear Recognition with Convolutional Neural Networks
DOI:
https://doi.org/10.19153/cleiej.25.2.8Keywords:
Ear recognition, CNN, Mask-RCNN, VGG16, VGGFace, transfer learning, domain adaptation, score-level fusionAbstract
Automatic recognition using ear images is an active area of interest within the biometrics community. Human ears are a stable and reliable source of information since they are not affected by facial expressions, do not suffer extreme change over time, are less prone to injuries, and are fully visible in mask-wearing scenarios. In addition, ear images can be passively captured from a distance, making it convenient when implementing surveillance and security applications. At the same time, deep learning-based methods have proven to be powerful techniques for unconstrained recognition. However, to truly benefit from deep learning techniques, it is necessary to count on a large-size variable set of samples to train and test networks. In this work, we built a new dataset using the VGGFace dataset, fine-tuned pre-train deep models, analyzed their sensitivity to different covariates in data, and explored the score-level fusion technique to improve overall recognition performance. Open-set and close-set experiments were performed using the proposed dataset and the challenging UERC dataset. Results show a significant improvement of around 9% when using a pre-trained face model over a general image recognition model; in addition, we achieve 4% better performance when fusing scores from both models.
Downloads
Published
Issue
Section
License
Copyright (c) 2022 Solange Ramos-Cooper, Guillermo Camara-Chavez
This work is licensed under a Creative Commons Attribution 4.0 International License.
CLEIej is supported by its home institution, CLEI, and by the contribution of the Latin American and international researchers community, and it does not apply any author charges whatsoever for submitting and publishing. Since its creation in 1998, all contents are made publicly accesibly. The current license being applied is a (CC)-BY license (effective October 2015; between 2011 and 2015 a (CC)-BY-NC license was used).