A Deep Learning Approach for Dog Face Verification and Recognition

Abstract

Recently, deep learning methods for biometrics identification have mainly focused on human face identification and have proven their efficiency. However, little research have been performed on animal biometrics identification. In this paper, a deep learning approach for dog face verification and recognition is proposed and evaluated. Due to the lack of available datasets and the complexity of dog face shapes this problem is harder than human identification. The first publicly available dataset is thus composed, and a deep convolutional neural network coupled with the triplet loss is trained on this dataset. The model is then evaluated on a verification problem, on a recognition problem and on clustering dog faces. For an open-set of 48 different dogs, it reaches an accuracy of 92% on a verification task and a rank-5 accuracy of 88% on a one-shot recognition task. The model can additionally cluster pictures of these unknown dogs. This work could push zoologists to further investigate these new kinds of techniques for animal identification or could help pet owners to find their lost animal. The code and the dataset of this project are publicly available (https://github.com/GuillaumeMougeot/DogFaceNet).