<p>Domain adaptive retrieval (DAR) is an important branch of transfer learning, that aims to address the distribution difference issue caused by domain differences between source and target domain. It achieves efficient image retrieval by transferring knowledge from the source domain to the target domain. However, existing methods still face three issues. (a) The irrelevant features of samples across domains may have a negative impact on the transferred knowledge. (b) Binary-valued labels possess limited semantic expressiveness relative to continuous-value representations. (c) Semantic preservation of hash codes incurs a large time overhead. To address the concerns raised above, we propose a novel DAR method based on Structure Center Fusion and Guidance Learning (SCFG). Specifically, SCFG first aligns the features between domains and extracts the relevant features from the different domains into the common subspace. Then, we fuse the structural information with the semantics of labels to construct category prototypes, which are used to replace labels to guide the learning of the subspace. In other words, we construct shared category prototypes in the subspace that are common to both domains, while also incorporating labels as membership information to enhance intra-class discriminability within the global structure. Finally, guided by pairwise orthogonal category prototypes, asymmetric semantic preservation is utilized to embed the category semantics into the hash codes, which generate highly discriminative hash codes, while improving computational efficiency by avoiding the use of paired cross-entropy losses. Experiments on multiple popular benchmark datasets demonstrate the low time overhead and effectiveness of SCFG.</p>