SPC: Selective Patches Consistency for Cross-Domain Few-Shot Learning

In recent years, Cross-Domain Few-Shot Learning (CD-FSL) has garnered significant attention due to its broad applicability across various domains. However, previous Few-Shot Learning (FSL) methods typically assume a consistent assignment between source and target domains, a condition that often leads to substantial performance degradation when applied to new domains. Furthermore, many existing approaches rely on only one or two distillation techniques, limiting their ability to capture local semantic information fully. An additional challenge arises from the imbalance in the loss function between tasks during training, which can cause overfitting to specific tasks or trigger instability. To overcome these challenges, we propose a novel CD-FSL approach, called SPC. Our method first extracts key local patches using the Acquisition ROI Module (AAR) and then integrates self-image distillation, cross-image distillation, and intra-image distillation (IID) to ensure consistency across local patches. This process significantly enhances the model’s ability to interpret critical regions. Moreover, we introduce a loss-driven multi-task dynamic weighting (LDW) mechanism, which adaptively adjusts the loss weights for each task during training. This dynamic adjustment balances task optimization and improves the model’s robustness and generalization. Experimental results on seven public cross-domain datasets, as well as one in-house medical dataset, demonstrate that our method outperforms existing approaches, improving model performance in cross-domain tasks. These results not only validate the effectiveness of our method but also highlight its potential for real-world applications.