Prototype Rectification with Region-Wise Foreground Enhancement for Few-Shot Classification

Few-Shot Classification is a challenging problem as it uses only very few labeled examples to assign labels for query samples. The Prototypical Network effectively addresses the issue by matching the nearest mean-based prototype for each query sample. However, the scarce labeled examples limit the ability of a model to represent the data distribution, which in turn biases the computed prototypes. In this paper, we propose a transductive bias diminishing method based on spatial similarity, which consists of a region-wise foreground enhancement (RFE) module and a prototype rectification (PR) module. RFE reconstructs the query sample’s features to emphasize the discriminative parts while preserving local region consistency. PR uses the difference in prediction scores between the nearest and second nearest prototypes of the enhanced query samples to rectify the prototypes by label propagation. We validate the effectiveness of our method with extensive experiments on four Few-Shot Classification datasets: miniImageNet, tieredImageNet, Stanford Dogs, and CUB. Our method achieves competitive performance across different settings and datasets.