Adaptive Deep Metric Learning with harmonized loss and nearest proxy alignment for low-dimensional representation

Deep Metric Learning (DML) plays a critical role in learning discriminative embedding spaces for tasks such as classification, retrieval, and dimensionality reduction. However, most existing methods rely on formulations such as pair-based or triplet-based losses, or proxy-based approximations that represent each class globally. They fail to provide effective discriminative information for hard-to-classify samples, lack the flexibility to adapt to inhomogeneous subclasses, and are highly affected by dominant samples. This makes the training process ineffective at distinguishing hard-to-classify samples or noise, resulting in suboptimal low-dimensional embeddings. To address these issues, we proposes an innovative Adaptive Deep Metric Learning framework, ADML_HNPA, which integrates a harmonized loss term that promotes separation of hard-to-classify samples and a nearest-proxy alignment term that improves intra-class compactness by emphasizing local structure. Specifically, we employ the log-ex approximation to define explicit objective for maximizing the margin between nearest neighbor within the same class and nearest neighbor from different class. We also define the objective for nearest proxy alignment. After that, we propose an adaptive parameter adjustment strategy to tune the parameters. Finally, extensive experiments are conducted on various datasets to demonstrate the superior of ADML_HNPA.