Current image-text retrieval methods typically rely on global feature alignment, neglecting the fine-grained matching of hierarchical semantic structures (e.g., local objects and phrasal relations) in images and texts, which limits retrieval accuracy in complex scenarios. To address this, we propose a Hierarchical Cross-Modal Alignment Network (HiCM) that achieves progressive semantic alignment from local to global by jointly modeling multi-scale features of both modalities. On the image side, we leverage the hierarchical architecture of Swin Transformer to extract multi-stage visual features, preserving hierarchical information from object parts to scenes. On the text side, we design a Dynamic Convolutional Multi-Scale (DCMS) module, which employs 1D convolutional kernels of varying sizes to capture word-level, phrase-level, and sentence-level semantic features. Furthermore, we introduce a hierarchical cross-modal alignment loss, enforcing image-text matching through a symmetric cross-entropy loss on global features while incorporating stage-wise contrastive losses to enhance local semantic alignment. Experiments on MSCOCO and Flickr30K demonstrate that HiCM significantly outperforms existing methods (e.g., achieving a 3.2% improvement in R@1 on MSCOCO). Ablation studies validate the effectiveness of text-side multi-scale modeling and stage-wise alignment. This work provides a hierarchical alignment paradigm for cross-modal understanding that balances efficiency and precision, particularly beneficial for real-world applications requiring fine-grained matching.