Automatic emotion recognition using electroencephalography (EEG) signals has garnered significant attention in recent years. While multi-source domain adaptation methods provide a promising framework for cross-subject emotion recognition, the distributional discrepancies among different source domains often result in negative transfer. To address these challenges, we propose a two-stage Dynamic Domain Selection and Pseudo-label Refinement (DDSPR) model. In the first stage, we introduce a novel Dynamic Domain Selection (DDS) module and an Agent Domain Adaptation Strategy (ADAS) to dynamically select and align source domains. In the second stage, a confidence-based pseudo-label correction strategy is employed to refine target domain labels and mitigate noise. We evaluate the proposed model through cross-subject experiments on the SEED and SEED-IV datasets, achieving accuracies of 91.50% ± 7.05 and 78.05% ± 13.56, respectively.The results demonstrate its effectiveness in emotion recognition performance.