<p>Unsupervised deep metric learning aims to construct a feature embedding space where similar samples are closely clustered while dissimilar samples are well separated, enabling effective similarity measurement without manual annotations. However, most existing methods still suffer from insufficient contextual information in the feature space and redundant background noise in images, which degrade metric performance. To address these issues, this paper proposes a context- and attention weighting-based unsupervised deep metric learning (CAUDML) method. Specifically, CAUDML jointly utilizes Gaussian similarity and cosine similarity to construct a composite similarity measure on the basis of the contextual distribution structure of the feature space, thereby guiding the model to learn more discriminative feature embeddings. Furthermore, a normalized attention weighting mechanism is introduced to perform weighted pooling on the final convolutional feature maps, which suppresses interference from background regions while enhancing the model’s representational capacity for target regions. Extensive experiments on three fine-grained image datasets validate the superiority and effectiveness of the proposed CAUDML method.</p>