Maker's marks and hallmarks on historical gold- and silverwork are crucial evidence for provenance research, cataloging, and object dating. However, identifying these marks manually is highly labor-intensive. Even in high-quality museum photographs, marks are diminutive, frequently worn, and often obscured by reflections or patina. To address this bottleneck, this paper presents and evaluates a computer-vision pipeline designed to automatically retrieve visually similar marks from challenging real-world datasets. We evaluate the impact of different localization strategies, including fully automated bounding-box prediction, alongside off-the-shelf and fine-tuned feature extraction networks. Our experiments demonstrate that combining localized cropping with metric-learning fine-tuning significantly improves retrieval accuracy, allowing the model to prioritize subtle iconographic features over background noise. These performance gains represent a crucial step toward scalable, AI-assisted identification tools for domain experts. By enabling the accurate retrieval of impressions struck by the same physical stamp, this work lays a foundation for advanced analyses of historical metalwork, including forgery detection and stamp abrasion studies.