FVSF: Fusion of Multi-Scale Visual and Enhanced Semantic Features for Few-Shot Learning

Few-Shot Learning (FSL) aims to recognize novel categories from sparse data but is fundamentally challenged by high intra-class variance and ambiguous feature representations. Existing methods are often limited by unreliable unimodal visual features or fail to bridge the modality gap due to inadequate semantic priors. To address these limitations, we introduce FVSF (Fusion of Visual and Semantic Features), a framework that constructs a highly discriminative embedding space by synergizing three key components. First, a Swin Transformerbased visual fusion module captures a rich hierarchy of visual features, from fine-grained textures to high-order semantics. Second, a Large Language Model (LLM)-driven pipeline generates descriptive, paragraph-level semantic representations for each category, resolving the ambiguity of conventional class labels. Third, a self-supervised contrastive learning strategy refines the embedding space to enhance intra-class compactness. Comprehensive experiments on standard FSL benchmarks, including MiniImageNet, CIFAR-FS, and FC100, demonstrate that FVSF significantly outperforms state-of-the-art methods, establishing a new performance benchmark. Our results confirm that the systematic integration of multimodal information provides a robust solution to the learning challenges posed by data scarcity in FSL.