Firearm brand classification using deep learning on cartridge case images.

Area of Science:

• Forensic Science
• Computer Science
• Materials Science

Background:

• Firearm identification relies on analyzing unique marks left on cartridge cases.
• Current ballistic examination systems use image matching but often miss firearm brand signatures.
• Identifying firearm brands can refine search spaces and enhance identification accuracy.

Purpose of the Study:

• To develop a deep learning approach for automated firearm brand classification using cartridge case surface topography.
• To improve the accuracy and efficiency of firearm identification in forensic ballistics.

Main Methods:

• Utilized the BALISTIKA system to generate high-resolution surface representations of over 350,000 cartridge cases from 21 major firearm brands.
• Applied normalized height maps and shape index transformation for feature extraction.
• Employed deep learning models (ResNet, Vision Transformer) and traditional machine learning (SVM, Random Forest) for classification.
• Mitigated class imbalance through oversampling minority classes with rotated samples, expanding the dataset to over a million samples.

Main Results:

• Deep learning models achieved superior performance, reaching up to 92% accuracy in firearm brand classification.
• The approach successfully classified cartridge cases from diverse firearm types, including handcrafted firearms and converted blank pistols (CBPs).
• Demonstrated the effectiveness of automated brand classification in prioritizing potential firearm matches.

Conclusions:

• Automated firearm brand classification using deep learning enhances the efficiency of forensic ballistics.
• This method allows forensic examiners to confidently narrow down comparisons to cartridge cases of the same brand.
• The proposed approach is expected to significantly reduce examination time and improve overall forensic investigation efficiency.