A novel and robust deep learning model for sibling firearm matching.

In this paper, our goal is to develop a novel and robust method for firearm sibling matching from cartridge-case head (in specific firing pin region) photometric stereo imagery. We employ a dataset of unprecedented scope compared to prior literature, incorporating genuine real-world case work, including cartridge cases fired from aged, worn, or deliberately modified firearms across different time periods, making it the most challenging setting. Methodologically, the study introduces a triplet-loss framework with semi-hard negative mining and an explicit margin term, combined with a pre-contrastive loss-an approach not previously explored in this domain. Performance is benchmarked against a deep model trained using pure contrastive loss as a most powerful method up to now used in this area. Our proposed model clearly outperforms with a large margin. Within first 20 candidates of models as siblings in test set, our model reaches to 90% of success while contrastive model 52%. Overall, the proposed study presents the strongest contribution in the literature on cartridge-case sibling matching, both in dataset realism and methodological advancement.