Abstract: Few-shot image classification suffers from limited feature representation capability due to the scarcity of training samples. This paper proposes a classification method that integrates a High-Low Frequency Adaptive Enhancement module and a Global Channel-Spatial Attention module, aiming to enhance multi-granularity frequency-domain features and enable multidimensional feature interaction. First, high-low frequency decomposition is applied to extract fine-grained texture and global structural information, and multi-scale convolutions are introduced in the high-frequency branch to capture semantic features under various receptive fields. Second, by combining channel attention, channel shuffle, and spatial attention mechanisms, the proposed method breaks the isolation between channels and promotes full fusion and interaction between channel and spatial features. Finally, a self-supervised contrastive learning strategy is adopted to introduce an auxiliary unlabeled task, effectively alleviating overfitting under limited data conditions. Compared with the baseline method SCL, the proposed method achieves consistent performance improvements on four standard few-shot classification benchmarks under both 1-shot and 5-shot settings, with gains of 1.59%/1.24% on miniImageNet, 1.83%/1.37% on tieredImageNet, 2.74%/2.16% on CIFAR-FS, and 1.91%/2.08% on FC100, demonstrating its effectiveness.