Attention-enhanced Relation Network for Few-shot Image Classification

Traditional deep learning models firmly rely on a large amount of labeled data during pre-training. Whereas it lacks generalization in the face of unfamiliar categories. Recently, few-shot learning is a hot topic in computer vision to classify unseen classes with limited labels. A representative approach is to extract features from the support and query sets, respectively, and compare similarities via metric learning. However, convolutional neural networks often focus only on a local region and ignore the global region, which severely reduces the accuracy of the matching. Specifically, in this paper, we pile lightweight attention-based blocks in the embedding module, which combines an adaptive kernel size 2D convolutional network with a cross-channel attention mechanism to encode multi-scale features and implicitly increase the receptive field. The SE-relation module chooses to construct learnable non-linear comparators to compare the relationship utilizing channel information. Finally, we show experimental results on standard few-shot testing benchmarks such as mini-ImageNet and tiered-ImageNet to demonstrate effectiveness.