WDSN: A Weighted Difference Scaling Network Based on the Prototypical Networks for Few-Shot Learning
Xiang Xu, Qili Chen · 2023
This paper presented a weighted difference scaling network (WDSN) based on the prototypical network. Firstly, for the intra-class distribution information, we designed a weighting module that assigned different weights to all samples in the same class according to the representativeness of the samples to generate more representative class prototypes. Secondly, we designed a differential scaling module aimed at further improving the classification performance of the prototype network by reducing intra-class differences and amplifying interclass differences. We conducted experiments on two benchmark datasets, and the experimental results showed that WDSN achieved significant progress compared to the prototypical network, and also outperformed some related methods in recent years.