Learning-based Ballistic Breech Face Impression Image Matching

Ballistic images of a cartridge case or bullet carry distinct "fingerprints" of the firearm, which is the foundation of widely used forensic examination in criminal investigations. In recent years, prior work has explored the effectiveness of correlation-based approaches in matching ballistic imagery. However, most of these studies focused on highly controlled situations and used relatively simple image processing techniques, without leveraging supervised learning schemes. Toward improving the matching accuracy, especially on operational data, we propose a learning-based approach to compute the similarity between two ballistic images with breech face impressions. Specifically, after a global alignment between the reference and probe images, we unroll them into the polar coordinate for robust feature extraction and global registration. A gentleBoost-based learning scheme selects an optimal set of local cells, each constituting a weak classifier using the cross-correlation function. Experimental results and comparison with state-of- the-art methods on the NIST database and a new operational database, OFL, obtained from Michigan State Forensics Laboratory demonstrate the viability of our approach.