In visible-infrared person re-identification, the limited number of training samples and the significant discrepancy between visible and infrared images pose key challenges in mining diverse cross-modal information and achieving feature alignment. An embedding enhancement network based on global learning expansion was proposed to address the challenges, which effectively integrated the strengths of both vision Transformer and CNN to jointly capture local details and global structural information from images. By generating diversified embedding features and introducing the feature map structural alignment mechanism, the model’s ability to model global structures was further enhanced and the modality gap between visible and infrared images was reduced, thereby enabling more comprehensive and discriminative feature representations to be learned. Experimental results demonstrated that, compared to the baseline model DEEN, the proposed method achieved improvements of 4.2 and 9.1 percentage points in Rank-1 accuracy under the full-search mode of the SYSU-MM01 dataset and the visible-light-to-infrared retrieval mode of the LLCM dataset, respectively. By aggregating global spatial information, the proposed method generated more discriminative feature embeddings, significantly boosting the network performance.