Abstract
Abstract
Ballistic examination systems rely on analyzing surface markings on cartridge cases to identify firearms used in criminal cases. However, current systems using traditional image processing techniques face limitations in accuracy and automation when handling large, diverse datasets. This thesis explores deep learning approaches to enhance both firearm brand classification and cartridge case matching in forensic ballistics.
We developed a firearm brand classification system using high-resolution surface height maps from the BALISTIKA system, focusing on 21 frequently encountered firearm brands representing 97% of forensic cases. We evaluated classical machine learning algorithms (Random Forest, SVM) and modern deep learning architectures (ResNet, Vision Transformers), leveraging shape index transformations and data oversampling with focal loss to address class imbalance. The ResNet-50 model achieved 91.6% accuracy, with visualization techniques confirming focus on forensically relevant features.
For cartridge case matching, we proposed a Siamese neural network to measure visual similarity between cases fired from the same firearm. Using comprehensive Controlled Sibling and Expert-Linked Sibling datasets, the model ranked evidence pools based on similarity to query images. The approach achieved a normalized AUC of 0.99, substantially outperforming the current BALISTIKA Generation 3 system (0.81 AUC). Occlusion sensitivity analysis revealed consistent focus on meaningful forensic features while ignoring irrelevant manufacturer markings.
These findings demonstrate that deep learning can significantly enhance forensic ballistic analysis by providing accurate, interpretable, and scalable solutions. The proposed methods represent substantial advancement toward automated ballistic identification systems suitable for real-world forensic deployment, potentially reducing expert workload while maintaining high standards required for criminal investigations.