Multi-scale kronecker-product relation networks for few-shot learning

Few-shot learning aims to train classifiers to learn new visual object categories from few training examples. Recently, metric-learning based methods have made promising progress. Relation Network is a metric-based method that uses simple convolutional neural networks to learn deep relationships between image features in order to recognize new objects. However, during the feature comparing phase, Relation Network is considered sensitive to the spatial positions of the compared objects. Moreover, it learns from only single-scale features which can lead to a poor generalization ability due to scale variation of the compared objects. To solve these problems, we intend to extend Relation Network to be position-aware and integrate multi-scale features for more robust metric learning and better generalization ability. In this paper, we propose a novel few-shot learning method called Multi-scale Kronecker-Product Relation Networks For Few-Shot Learning (MsK-PRN). Our method combines feature maps with spatial correlation maps generated from a Kronecker-product module to capture position-wise correlations between the compared features and then feeds them to a relation network module, which captures similarities between the combined features in a multi-scale manner. Extensive experiments demonstrate that the proposed method outperforms the related state-of-the-art methods on popular few-shot learning datasets. Particularly, MsKPRN has improved the accuracy of Relation Network from 50.44 to 57.02 and from 65.63 to 72.06 on 5-way 1-shot and 5-shot scenarios, respectively. Our code will be available on: https://github.com/mouniraziz/MsKPRN.