In data classification and retrieval tasks, the rationality of similarity measurement directly determines task performance. Traditional similarity measures, such as Euclidean distance and cosine similarity, treat all dimensions equally and fail to capture the discriminative power of different features, thereby limiting effectiveness. Although deep learning–based metric learning methods can automatically extract features, they often overlook local structural information and rely heavily on label supervision. To address these issues, this paper proposes a deep metric learning approach based on local structural relationships. Specifically, a transfer architecture is introduced to propagate neighborhood relationships and preserve local structures, while a neighborhood matrix is employed to replace labels for unsupervised training. Extensive experimental results verify the effectiveness of our local structure-aware approach.