PFCC-Net: A Polarized Fusion Cross-Correlation Network for efficient and high-quality embedding in few-shot classification

Abstract

Constructing an effective embedding space is critical for metric-based few-shot learning. However, existing methods often struggle to balance embedding quality with computational cost. In this paper, we propose the Polarized Fusion Cross-Correlation Network (PFCC-Net), which leverages inter-image relationships to generate high-quality embeddings. The high-quality embedding depends on two core modules: the Parallel Path Fusion (PPF) module, designed to enhance multi-level embeddings while preserving rich semantic information, and the Polarized Hybrid Cross-Attention (PHCA) module, which performs polarization filtering on spatial and channel branches to extract salient features. Experimental results demonstrate that PFCC-Net achieves competitive results on four widely used few-shot classification benchmarks: miniImageNet, tieredImageNet, CUB-200-2011, and CIFAR-FS, while significantly reducing computational costs compared to ViT-based methods.