Multi-branch semantic alignment for few-shot image classification

The remarkable progress of deep learning in computer vision has significantly stimulated research interest in few-shot image classification. This field aims to transfer knowledge from previous experiences to recognize new concepts with limited samples. However, most existing approaches primarily concentrate on aligning semantic information at high-level features, neglecting the importance of middle-level or low-level feature representations. In this paper, we propose a novel approach called Multi-Branch Semantic Alignment (MBSA) for few-shot image classification, with the objective of investigating the role of multi-level features. Instead of using standard convolutional layers, we employ diverse convolutional layers to generate enhanced representations in each branch. These representations are then utilized by a dense classifier, which is supervised by a powerful guidance mechanism to incorporate semantic information into their spatial locations. During the inference stage, the multi-branch semantic alignment is designed to align multi-level features between query images and support images. This alignment process effectively establishes semantic correspondences between representations at different levels, thereby enhancing the ability to recognize novel categories. Comprehensive experiments are conducted on various few-shot benchmarks to demonstrate the superiority of our approach compared to those of several previous approaches, and ablation studies are performed to analyze the impact of different components.