Cross-Domain Few-Shot Classification based on Lightweight Res2Net and Flexible GNN

摘要

Cross-Domain Few-Shot Classification aims to recognize new categories from unseen domains while each category has only a few support examples. But existing networks cannot be effectively applied to cross-domain scenario. To solve this problem, in this paper, we propose two new strategies, respectively for the encoder and the metric function of metric-based network: First, we propose a precise metric function named FGNN(Flexible GNN) to better measure the distance between images whether labeled or unlabeled; Second, based on the idea of multi-scale representation, we build a new hierarchical residual-like block which is applicable to lightweight ResNet structures such as ResNet-10. The constructed network—LR2Net(Lightweight Res2Net), performs much better than ResNet and provides a new scale-based strategy to constantly increase precision. Various feature encoders combined with metric function GNN or FGNN are verified through a lot of contrast experiments using leave-one-out setting on four datasets—CUB, Cars, Places and Plantae. As a result, the highest average precision of our combined networks achieves up to 2.22% and 2.26% improvement compared to the state-of-art under the 5-way 1-shot and 5-way 5-shot cross-domain classification.