One-Shot Image Learning Using Test-Time Augmentation

EasyChair Preprint 7185

Abstract

Modern image recognition systems require a large amount of training data. In contrast, humans can learn the concept of new classes from only one or a few image examples. A machine learning problem with only a few training samples is called few-shot learning and is a key challenge in the image recognition field. In this paper, we address one-shot learning, which is a type of few-shot learning in which there is one training sample per class. We propose a one-shot learning method based on metric learning that is characterized by data augmentation of a test target along with the training samples. Experimental results demonstrate that expanding both training samples and test target is effective in terms of improving accuracy. On a benchmark dataset, the accuracy improvement by the proposed method is 2.55 percent points, while the improvement by usual data augmentation which expands the training samples is 1.31 percent points. Although the proposed method is very simple, it achieves accuracy that is comparable or superior to some of existing methods.

Keyphrases: image classification, one-shot learning, test-time augmentation

@booklet{EasyChair:7185,
  author    = {Keiichi Yamada and Susumu Matsumi},
  title     = {One-Shot Image Learning Using Test-Time Augmentation},
  howpublished = {EasyChair Preprint 7185},
  year      = {EasyChair, 2021}}