In deep metric learning, the expressiveness of proxy points plays a crucial role in determining the quality of class representations. However, in real-world datasets, samples from the same class often present multiple sub-cluster distributions due to variations in local features such as pose and texture. A single proxy is usually insufficient to capture these diverse intra-class variations, thus limiting the method's discriminative ability. To tackle this issue, we propose a deep metric learning method based on dynamic multi-proxy (DMP-DML). Specifically, we introduce a memory bank to store samples that are poorly represented by the current proxies during training. New proxies are then dynamically estimated from high-density regions within the memory bank, ensuring that they effectively capture key feature distributions of the class. Furthermore, we design an adaptive proxy selection strategy that evaluates the importance of candidate proxies to filter redundancy. Finally, a combined proxy-sample and proxy-proxy loss function is proposed, which encourages tight clustering of intra-class samples while pushing inter-class proxies farther apart. Experimental results on three standard benchmark datasets show that DMP-DML outperforms existing deep metric learning methods, demonstrating its effectiveness in capturing intra-class diversity and improving the discriminative power of the embedding space.