A Multi-perspective Squeeze Excitation Classifier Based on Vision Transformer for Few Shot Image Classification

Few-shot image classification is a task that uses a small number of labeled samples to train a model to complete the classification task. Most few-shot image classification methods use small CNN-based models due to its good performance under supervised learning. However, small CNN-based models have performance bottlenecks under self-supervised learning with a large amount of unlabeled data. So we propose a model based on ViT for few-shot image classification. We propose a method combining Mask Image Modeling self-supervised learning and cross-architecture knowledge distillation to improve ViT. For few-shot image classification task, we propose a multi-perspective squeeze-excitation projector that is able to exploits the mutual information between samples in different perspectives, and aggregate in-class samples and discretize out-of-class samples. Finally, we construct a classifier based on it. Experimental results on Mini-ImageNet and Tiered-ImageNet show that our model achieves an average of 2% improvement over the previous state-of-the-art.