A Fast Classification Method for Few‐Shot Images Based on Data Augmentation

ABSTRACT

Existing data augmentation methods for few‐shot image classification often suffer from high model complexity and prolonged inference time. To address these limitations, this paper proposes a tensor feature generator that directly operates in the feature space to augment few‐shot image data by synthesising new tensor features, without the need to convert between tensor and vector representations. Based on this generator, we develop a fast classification network (TF‐FCN) for few‐shot images. The architecture integrates a residual network for tensor feature extraction and the proposed generator for feature‐level augmentation, enabling effective classifier training with low inference latency. Extensive experiments on mini‐ImageNet, CUB, and CIFAR‐FS benchmarks demonstrate that TF‐FCN achieves competitive or superior classification accuracy compared to recent state‐of‐the‐art data augmentation methods. Moreover, when using ResNet18 as the backbone, TF‐FCN reduces inference time by up to 48% as the number of generated tensor features increases, while maintaining robust performance across multiple classifiers. These results underscore the effectiveness and efficiency of TF‐FCN for few‐shot image classification.