Fired bullet matching candidate retrieval algorithm based on Siamese neural network

In gun-related crime investigations, analyzing fired bullets is critical. Traditional examination via comparison microscopes is labor-intensive and time-consuming, hindering investigation efficiency. To address these issues, this study uses a Siamese neural network (SNN) to support bullet analysis through similarity-based candidate retrieval. A dataset of 17,870 bullets from 8935 firearms (two bullets per firearm) was built; their surface topographies were scanned into 2D images via the BalScan system, preprocessed, and split into training/testing sets. The SNN was trained on training data, and instance retrieval was used on testing data to retrieve similarity-based candidates for unidentified bullets. Results showed the method's effectiveness: against a 600-image gallery, top 1, top 5, and top 10 retrieval accuracies reached 80.2%, 93.4%, and 97.3%, respectively. Compared with previous studies with limited firearms and more bullets per firearm, this work expanded firearm samples and increased task complexity. The SNN-based retrieval system can expedite the process of narrowing down potential firearm matches for bullets, supporting examiners in ballistic examinations and aiding forensic investigations.