In cloud environments, encrypted image retrieval can effectively protect data privacy; however, limited feature expressiveness and weakened visual semantics in the ciphertext domain remain major challenges that constrain retrieval accuracy. To address these issues, an encrypted image retrieval method that integrates dynamic feature representation with cross-modal semantic co-constraints was proposed. First, a non-symmetric cross-attention-based local-global dynamic feature fusion network was developed to model the complementary relationships between fine-grained textures and global semantics, thereby enhancing the discriminability of visual features. Second, a visual-text semantic constrained cross-modal hashing framework was designed to project fused visual features and textual features into a unified hashing space, where textual semantics provide weak supervisory guidance to reduce visual ambiguity and improve cross-modal consistency. Third, the BFV(Brakerski-Fan-Vercauteren) homomorphic encryption scheme was incorporated to enable similarity computation directly in the encrypted domain, ensuring the privacy of both feature representations and retrieval results. Experiments conducted on Olivia-2688, GHIM-10k, Caltech-101, and NWPU-RESISC45 demonstrate that the proposed dynamic fusion module improves mAP(mean average precision) by 14.01%~15.23%, while the cross-modal semantic constraint yields an additional 4.56%~9.93% performance gain. Furthermore, when the polynomial ring degree is set to 8 192, ciphertext-domain similarity computation is fully consistent with its plaintext counterpart and exhibits high computational efficiency. The results confirm that the proposed method achieves highly discriminative encrypted feature representation and supports high-precision encrypted image retrieval in complex cloud environments.