Cross-Domain Few-Shot Classification through Diversified Feature Transformation Layers

Abstract

The purpose of the few-shot classification is to classify new categories, and each category contains few labeled samples. The currently popular cross-domain few-shot classification uses a feature transformation layer to transform features to achieve the feature enhancement, so as to simulate various feature distributions in different domains during the training process. However, due to the large differences in the distribution of cross-domain features, a single feature transformation layer cannot perform multiple feature transformations. To obtain the change of the feature distribution in different domains, a diversified feature transformation is proposed based on the original feature transformation layer to solve the metric-based cross-domain few-shot classification problem Simulation results are obtained based on these five datasets commonly used in few-shot classification: mini-ImageNet, CUB, Cars, Places and Plantae. The simulation results show that the proposed diversified feature transformation layer can achieve good results in the metric-based model.