Few-Shot Learning with Improved Local Representations Via Bias Rectify Module

Chao Dong,Qi Ye,Wenchao Meng,Kaixiang Yang
DOI: https://doi.org/10.1109/icassp43922.2022.9746156
2022-01-01
Abstract:Recent approaches based on metric learning have achieved great progress in few-shot learning. However, most of them are limited to image-level representation manners, which fail to properly deal with the intra-class variations and spatial knowledge and thus produce undesirable performance. In this paper, we propose a Deep Bias Rectify Network (DBRN) to fully exploit the spatial information that exists in the structure of the feature representations. We first employ a bias rectify module, which is able to focus on the features that are more discriminative for classification by giving different weights, to alleviate the adverse impact caused by the intra-class variations. To make full use of the training data, we design a prototype augment mechanism that can make the prototypes generated from the support set to be more representative. To validate the effectiveness of our method, we conducted extensive experiments on various popular few-shot classification benchmarks and our methods can outperform state-of-the-art methods.
What problem does this paper attempt to address?