Style-Guided Source Data Augmentation and Target Feature Optimization for Cross-Domain Few-Shot Image Classification

In Cross-Domain Few-Shot Learning (CD-FSL), models are required to identify novel classes while addressing domain discrepancies caused by visual style variations. Simple style transformations often fail to extend beyond the source domain’s distribution, and unrepresentative support samples in the target task may lead to ambiguous or biased decision boundaries. To address these challenges, a Style-Guided Source Data Augmentation and Target Feature Optimization (SSDATFO) approach is proposed. Specifically, Style-Guided Source Data Augmentation is introduced, employing Style Transformation and Source Data Augmentation techniques to create more challenging source data, thereby expanding the source domain’s style distribution. Target Feature Optimization is subsequently introduced, comprising two distinct modules. The Domain Attention Shift Transformation enhances low-magnitude feature channels, thereby reactivating target domain feature channels previously overlooked by the source domain-trained feature extractor. Additionally, the Task Category Differentiation Enhancement Transformation calibrates the features of support samples and eliminates the commonality component along both the task-specific and inter-class commonality directions for all features within the novel task, thereby acquiring more discriminative features. Extensive experiments on eight distinct target datasets demonstrate the efficacy of the proposed method, while comprehensive ablation studies and detailed visualization experiments elucidate its nuanced and compelling aspects.