Global- and local-aware feature augmentation with semantic orthogonality for few-shot image classification

As for few-shot image classification, recently, some works revisit the standard transfer learning paradigm, i.e., pre-training and fine-tuning, and have achieved some success. However, we find that this kind of methods heavily relies on a naive image-level data augmentation (e.g., cropping and flipping) at the fine-tuning stage, which will easily suffer from the overfitting problem because of the limited-data regime. To tackle this issue, in this paper, we attempt to perform a novel feature-level semantic augmentation at the fine-tuning stage and propose a Global- and Local-aware Feature Augmentation method (GLFA) from both the channel- and spatial-wise perspectives. In addition, at the pre-training stage, we further propose a Semantic Orthogonal Learning Framework (SOLF) to make the learned feature channels more independently, orthogonal and diverse. Extensive experiments demonstrate that the proposed method can obtain significant performance improvements over the state of the arts. Code is available at https://github.com/onlyyao/GLFA-SOLF.