Enhancement of Few-shot Image Classification Using Eigenimages

In this paper, we propose an auxiliary loss function called an eigen loss to reduce the overfitting of few-shot learning algorithms. The proposed loss function predicts the class of unlabeled query images by measuring the similarity between the query image and reconstructed image constructed from the eigenimages of the support data. The eigen loss is used in a linearly combined form with the existing loss function of few-shot learning models. Experimental results of the eigen loss applied to representative few-shot learning models on widely used datasets (i.e., MiniImageNet, CUB, and TieredImageNet) show that the proposed method yields notable improvements in terms of classification accuracy.