Learning to Calibrate Prototypes for Few-Shot Image Classification

Chenchen Liang, Chenyi Jiang, Shidong Wang, Haofeng Zhang

Abstract: Few-shot learning (FSL) aims to generalise the model to novel classes by using a limited amount of discriminative samples (a.k.a., prototypes). With few labelled samples, there is much uncertainty and randomness in the data, which makes it more difficult for the model to learn the complete underlying patterns. This paper proposes a Discriminative Property Calibration Network (DPCNet) to enable model building in linear space with robust separability. Concretely, the property features of samples are extracted to facilitate filtering out low-informative key points at the instance level, and then the key points are further refined from the perspective of property features to retain those dimensions that contain the most relevant properties. Furthermore, the discriminative key properties are re-weighted by accounting for the correlation between images, thus forcing the model to focus more on the key property information. Moreover, a new margin algorithm is proposed to optimise the data distribution of features by dynamically adjusting the distance between classes. We conduct extensive experiments on four datasets, i.e., miniImageNet, tiredImagenet, CUB-200-2011 and CIFAR-FS, achieving the accuracies of 67.96%, 72.57%, 79.6% and 74.56%, respectively, on the 5-way 1-shot setting, and the same very competitive performance on the 5-way 5-shot setting. The proposed method can well extract the most relevant and discriminative properties, the re-weighted features further emphasise the discrimination and the dynamic margin algorithm enhances the stability and generalisation ability. The proposed method achieves the state-of-the-art performance, and it will have meaningful inspiration for future works.

External IDs:dblp:journals/cogcom/LiangJWZ25