Knowledge-Guided Distribution Alignment for Cross-Domain Few-Shot Learning

Cross-domain few-shot learning (CD-FSL) aims to recognize novel categories with minimal labeled samples in target domains that differ from source domains. However, difficulty in obtaining valid domain bias guidance leads to negative transfer challenges because the target domain samples are unknown during source domain training. Inspired by human reliance on prior knowledge when adapting to new domains, we propose a knowledge-guided distribution alignment network (KDANet). In contrast to earlier CD-FSL approaches that primarily focus on visual alignment alone, KDANet incorporates textual priors in both the training and adaptation stages, thereby enhancing domain transferability. Specifically, KDANet integrates textual priors as knowledge guidance for visual representation learning during pretraining in the source domain with sufficient samples to establish an initial learning foundation. To tackle the scarcity of target domain samples, available samples and construct pseudo-episodes are expanded through critical region detection. Leveraging the pretrained model and pseudo-episodes, a two-stage progressive finetuning method is employed to refine feature extraction and calibrate prototypes for target domain tasks, with prior knowledge guiding the learning process continuously. Moreover, adaptive distribution alignment is proposed throughout cross-domain training and finetuning to suppress distribution bias interference by utilizing multi-source domain alignment and triplet supervision. Quantitative and qualitative experiments demonstrate the superior performance of proposed method, particularly in challenging 1-shot tasks. Under the CD-FSL benchmark, proposed method achieves an average accuracy improvement of 3% across all target domains, outperforming state-of-the-art methods. Code is available in https://github.com/HHU-JialeChen/KDANet.git

Recommended citation: Jiale Chen, Feng Xu, Xin Lyu, Tao Zeng, Xin Li, and Shangjing Chen. Knowledge-Guided Distribution Alignment for Cross-Domain Few-Shot Learning. Knowledge-Based Systems, 318:114316, 2025.
Download Paper