Most cross-modal hashing methods use binary matrix to represent the degree of correlation, which results in high-level semantic information cannot be captured in multi-label data, and those methods ignore maintaining the semantic structure and the discrimination of the data features. Therefore, a cross-modal retrieval algorithm named ML-SDH (Multi-Level Semantics Discriminative guided Hashing) was proposed. In the algorithm, multi-level semantic similarity matrix was used to discover the deeply correlated information in the cross-modal data, and equally guided cross-modal hashing was used to express the correlations in the semantic structure and discriminative classification. As the result, not only the purpose of encoding multi-label data of high-level semantic information was achieved, but also the distinguishability and semantic similarity of the final learned hash codes were ensured by the constructed multi-level semantic structure. On NUS-WIDE dataset, with the hash code length of 32 bit, the mean Average Precision (mAP) of the proposed algorithm in two retrieval tasks is 19.48,14.50,1.95 percentage points and 16.32,11.82,2.08 percentage points higher than those of DCMH (Deep Cross-Modal Hashing), PRDH (Pairwise Relationship guided Deep Hashing) and EGDH (Equally-Guided Discriminative Hashing) algorithms respectively.