Low-resolution few-shot learning via multi-space knowledge distillation

Abstract

Existing few-shot classification models usually rely on limited known support images to form class centers, and classify query images based on the distance between their embedding and the class centers. However, these models assume that the query image is high-resolution (HR), and thus suffer from significant performance degradation when applied to low-resolution (LR) images. Due to the lack of discriminative information in LR images, there is a noticeable discrepancy between the embeddings of LR query images and the class centers formed by HR support images. To address this issue, we first formulate the problem of Low-Resolution Few-Shot Learning (LRFSL), where the support images are HR while the query images are only available in LR. Then, we propose an end-to-end pipeline that leverages mutual learning between a super-resolution (SR) network and a few-shot classification network. To further reduce the domain discrepancy between the embeddings of the SR images and HR class centers, we introduce a multi-space knowledge distillation strategy that aims to transfer pixel-level, feature-level, and logit-level knowledge of the HR domain to the SR domain. We conduct extensive experiments on classic few-shot datasets: miniImageNet, tieredImageNet, and the fine-grained few-shot dataset CUB. Experimental results show that our method can handle few-shot classification with LR input, and achieve performance that is almost comparable to using HR images as input. Specifically, our method achieves an average accuracy improvement of 26.47% with the Meta-Baseline Model and 7.44% with the Meta DeepBDC Model across all datasets compared to LR Query.