MTVS-Few: Prototype Alignment with Multi-View Semantic Information for Few-Shot Learning

Few-Shot Learning aims to establish comprehensive class conceptual understanding and achieve accurate recognition from limited samples. Recent researches focus on using multimodal information, particularly by incorporating text to compensate for the discriminative feature deficiencies in image modality. However, relying on low-quality text such as class name and implicit attributes still fails to bridge the representation gap between images and prototypes. In this paper, we propose a framework that generates high-quality multi-view textual descriptions as complementary knowledge through progressive task-driven prompt from large language models. To select appropriate textual semantics from the image perspective, we introduce the attention mechanism to adaptively generate the proper textual representation. Unlike previous methods, our approach employs a KAN-based alignment network rather than MLP for inter-modal semantic fusion and representative prototype reconstruction within Yeo-Johnson distribution calibration to enhance capabilities of Few-Shot Learning. Experimental results across various benchmark datasets demonstrate the effectiveness of our model (MTVS-Few), surpassing many state-of-the-art Few-Shot Learning methods.