Firearm brand classification using deep learning on cartridge case images

Abstract

When a firearm is discharged, it leaves characteristic marks on the cartridge case, which are analyzed in forensic ballistics to identify the firearm. Conventional ballistic examination systems rely on high-quality images of cartridge cases and bullets, scanning databases to generate ranked candidate lists based on similarity scores. However, these systems often overlook the distinctive signatures of the firearm brand, which could refine search spaces and improve identification accuracy. In this study, we propose a deep learning-based approach leveraging normalized height maps and shape index transformation of cartridge cases for firearm brand classification. Using the BALISTIKA system, we generated high-resolution surface representations from over 350,000 cartridge cases representing the most populous 21 firearm brands, representing 97% of firearms encountered in criminal cases in Türkiye, including handcrafted firearms and converted blank pistols (CBPs). By oversampling the minority classes in the dataset using rotated samples, we expanded it to over a million samples and mitigated class imbalance. We evaluated both traditional machine learning (SVM, Random Forest) and deep learning models (ResNet, Vision Transformer), with deep learning approaches achieving superior performance of up to 92% accuracy. These findings demonstrate that automated firearm brand classification enables forensic examiners to confidently prioritize cartridge cases from the same brand during ballistic comparisons. This approach is expected to substantially reduce examination time and enhance the efficiency of forensic investigations.