Few-shot learning aims to train classifiers with limited samples for novel object recognition, facing key challenges in feature extraction robustness and discriminative representation. To address these issues, we propose a Median-Enhanced Multi-Scale Adaptive Network. Firstly, an adaptive fusion convolution module with deformable kernels is designed to capture spatially transformed features, improving cross-domain adaptability. Next, a median-enhanced attention mechanism integrates median filtering with channel attention, effectively suppressing feature noise and outliers while highlighting discriminative patterns. Finally, we develop a hierarchical metric learning framework that combines multi-scale feature representations with learnable similarity metrics. Experimental results demonstrate that the proposed method outperforms state-of-the-art approaches, achieving accuracy gains of 1.27% (1-shot) and 1.12% (5-shot) on Mini-ImageNet, 1.76%/1.52% on Tiered-ImageNet, and 2.28%/2.21% on CUB, compared to the SetFeat model.
Yang et al. (Mon,) studied this question.