Among cross-modal retrieval methods, hashing methods are one of the important approaches. In existing hashing methods, similarity matrices are often predefined based on fixed label relationships, ignoring semantic correlations between labels and inherent data features, leading to inadequate representation of inter-modal similarity relationships. We propose a new hashing method that designs a dynamically updated similarity matrix to adapt to the real intra- and inter-modal similarity structures. By adaptively learning latent label correlations and dynamically optimizing the similarity matrix, the method optimizes the representation of similarity relationships via spatial consistency constraints to maintain consistency between data features and labels as much as possible. Experiments on three datasets show that this method significantly outperforms existing baseline methods across different hash code lengths, verifying the effectiveness and advancement of adaptive modeling of similarity relationships in cross-modal task.