메타학습 기반 소수샷 반려견 얼굴 식별

초록

As the population of companion animals increases and the problem of lost or abandoned dogs intensifies, research in dog face recognition has focused on improving performance on individual identification datasets. However, conventional studies have shown limitations in addressing the fundamental challenge of few-shot environments, where the number of images per individual is critically low. To address this issue, this study proposes a meta-learning-based few-shot face recognition system for rapidly identifying individual dogs from a small number of images. We conducted comparative experiments on the DogFaceNet dataset using two meta-learning techniques: Prototypical Networks and Meta-DeepBDC. The results show that Meta-DeepBDC achieved a high classification accuracy of 64.01% in a 1-shot setting with a ResNet-12 backbone. This study marks the first application of meta-learning to the field of dog face recognition and is significant for demonstrating that individual identification performance can be effectively enhanced even in data-scarce environments.

키워드