Group equivariant learning for few-shot image classification

Abstract

Few-shot learning, as an effective approach to solve image classification problems in data-scarce scenarios, has made significant progress in recent years, with numerous methods emerging. These methods typically use convolutional neural networks (CNNs) as feature extractors and classify other data based on the features of a small number of labeled samples. The reason CNNs have become the preferred method for image processing tasks is primarily due to their translational equivariance. However, conventional CNNs lack inherent mechanisms to handle other symmetry transformations (such as rotation and reflection), resulting in reduced classification performance of the model, especially in few-shot scenarios. To address this problem, we leverage the advantages of group convolutions in handling broader symmetric transformations, integrating them into few-shot learning tasks, and accordingly propose a group-equivariant prototypical learning network. This method maps samples into the group space via a group convolution module, enhancing the model’s ability to handle various symmetry transformations present in classification targets within images, thereby improving its feature representation capability. Additionally, we designed a new contrastive loss that can naturally be co-optimized with cross-entropy loss, guiding the model to learn a highly discriminative group feature space. The experimental results on the miniImageNet, CIFAR-FS, and CUB-200 datasets show that the GEPL method significantly improves classification performance, thus verifying the effectiveness of our method.