PMGCN:Preserving Measuring Mapping Prototype Graph Calibration Network for Few-Shot Learning

The aim of few-shot classification is to learn the discriminative features from a limited sample of labeled data. Due to the few number of labeled data, the model can not learn the discriminative knowledge, and can usually become overfitted and achieve a bad generalization effect. In previous work, the training data is not mentioned after the model is trained. However, considering the few-shot learning paradigm, we try to use the training data in more places when the data set itself is not sufficient. In this paper, we propose a new method Preservation measuring Mapping Prototype Graph Calibration Network(PMGCN) based on the prototypical networks to fully utilize the training dataset to obtain a more accurate prototype. In order to get more discriminative feature, we first use a Convolutional Block Attention Module(CBAM) to focus on the key parts of the data and ignore the noisy parts, which is a simple yet effective attention module. Then, in order to fully utilize the limited labeled data, we propose to use a graph convolutional network(GCN) combined with preserving measuring mapping(PMM) in ergodic theory to select some similar base class prototypes to calibrate the new prototypes to obtain better generalization and make the model more interpretable. We validate our model on mimiimagenet, tieredimagenet and cub. The performance of the proposed method is improved by \(2.24\%\) and 2.86%, \(1.35\%\) and \(2.19\%\) , \(1.42\%\) and \(0.82\%\) separately.