<p>Cross-domain few-shot object detection faces challenges of limited target domain samples and significant domain heterogeneity, which severely tests traditional pseudo-label-based transfer methods. Due to limited single-machine computational resources, traditional models struggle to handle large-scale cross-domain data, failing to capture the complex nonlinear structures and deep semantic relationships of the target domain, leading to increased domain differences, amplified prediction uncertainty, and reduced knowledge transfer efficiency. To address these issues, we propose a pseudo-label refinement and uncertainty-guided alignment optimization for cross-domain few-shot object detection. First, to address the issue of limited samples, we designed a dual-branch joint enhancement module to improve the expressive power of limited data, thereby acquiring more discriminative information. To improve the quality of pseudo-labels, we introduce a feedback-driven progressive pseudo-label optimization strategy. This approach iteratively improves pseudo-label quality through collaboration between a teacher and a student model. Second, to handle low-confidence pseudo-labels, we propose a soft-weighted balancing strategy. This method adjusts the weight of low-confidence pseudo-labels, reducing noise while preserving valuable weak supervision signals. To mitigate the uncertainty in pseudo-label generation, we also introduce an uncertainty alignment loss. Experimental results on six datasets demonstrate that our method outperforms baseline approaches, significantly improving pseudo-label quality and enhancing cross-domain transfer learning performance.</p>