Local descriptor-based multi-prototype network for few-shot Learning

作者：

Highlights：

• We present a simple method to effectively approximate the underlying distribution of a class by using multiple prototype learning.

• We use the channel squeeze and spatial excitation mechanism to selectively emphasise informative local descriptors and suppress less useful ones.

• Our method outperforms some metric-based FSL methods and achieves competitive performance over other meta-based methods on multiple benchmarks.

摘要

•We present a simple method to effectively approximate the underlying distribution of a class by using multiple prototype learning.•We use the channel squeeze and spatial excitation mechanism to selectively emphasise informative local descriptors and suppress less useful ones.•Our method outperforms some metric-based FSL methods and achieves competitive performance over other meta-based methods on multiple benchmarks.

论文关键词：Few-shot learning,Image classification,Local descriptors,Multiple prototypes,End-to-end learning

论文评审过程：Received 8 February 2020, Revised 17 January 2021, Accepted 7 March 2021, Available online 10 March 2021, Version of Record 27 March 2021.

论文官网地址：https://doi.org/10.1016/j.patcog.2021.107935