Bias-in-debias-out: Hierarchical channel-spatial bias calibration for cross-domain few-shot classification

Abstract

The core challenge of cross-domain few-shot learning (CD-FSL) stems from models’ inability to generalize source-domain inductive biases to target domains under significant distribution shifts. While existing methods predominantly employ strategies like auxiliary target data adaptation, feature disentanglement, or metric space alignment, they overlook two inherent biases entrenched during source-domain training: (1) channel-wise dependency on source-specific feature patterns and (2) spatial-wise preference for source-typical structures, both of which hinder cross-domain transfer. We propose the first unified Channel-Spatial Dual-dimensional Bias Calibration (CSDBC) framework to systematically address these biases through progressive dilution, recomposition, and alignment. Our approach integrates three key innovations: (1) a parameter-free Static Base-class Bias Dilution (SBBD) module that dilutes source-specific channel-spatial biases through layer-wise and point-wise modulation, effectively suppressing overfitting to source-specific patterns; (2) a Dynamic Novel-class Bias Recomposition (DNBR) module that generates target-adaptive channel-spatial soft masks via meta-optimized lightweight depthwise separable convolutions, enabling target-domain channel reweighting and spatial preference adjustment; and (3) a Novel-class Cross-image Semantic Alignment (NCSA) module that establishes channel correlations and spatial correspondences between support-query pairs, significantly enhancing both discriminability and semantic consistency of target-domain features. Extensive experiments across eight CD-FSL benchmarks demonstrate consistent improvements, outperforming SOTA methods by 1.35 % (5-way 1-shot) and 2.00 % (5-way 5-shot) in average accuracy under varying domain shifts.