Image–text retrieval serves as a pivotal bridge between vision and language, playing a critical role in achieving high-quality semantic alignment across modalities. To enhance the fine-grained discriminative capacity of semantic features in cross-modal retrieval, this study proposes a multi-modal alignment framework that integrates intra-modal fine-grained local semantics with inter-modal global semantic correlations. Original features of images and texts are first extracted, followed by region-level relational reasoning through Graph Attention Networks (GAT) to obtain relation-enhanced local features. An intra-modal attention mechanism is then employed to facilitate semantic interactions among samples within the same modality, thereby capturing comprehensive intra-modal associations and producing semantically enriched embeddings for both images and texts. Finally, a triangular constraint–enhanced triplet loss function is designed to optimize the alignment process. Experimental results on the MS-COCO dataset (1K test set) demonstrate improvements of 6.4% and 1.3% in image retrieval and text retrieval tasks, respectively, in terms of Recall@1, confirming the effectiveness of the proposed method.