Advanced deep learning for automatic classification of fired bullets from standard-issue firearms.

Gun violence continues to be a significant global issue, causing countless innocent lives to be lost each year. This study explores deep learning for automated fired bullet marking classification. To address inconsistent results from examiners' subjective same-source judgment, we automate classification to boost forensic firearm examination accuracy and reduce subjectivity. We collected a dataset of 6000 fired bullets from six types of standard-issue firearms commonly used by Chinese law enforcement agencies. Panoramic images of the lateral surfaces of the bullets were captured using the BalScan system. To create diverse and informative inputs for our deep learning model, we employed three distinct image preprocessing strategies: panoramic imaging, land engraved area (LEA) segmentation (the area with striations and grooves created by the gun barrel's rifling), and line segmentation. Then, we fine-tuned the advanced pre-trained ResNet50 network on this dataset, specifically designed for image classification tasks. Our experiments demonstrated the effectiveness of our approach, achieving high classification accuracy across different firearm types. Notably, the LEA segmentation strategy outperformed the other methods, highlighting the importance of focusing on specific regions of interest for accurate classification. To specify, our algorithm with the LEA segmentation strategy achieves a classification accuracy of 97.2% for six types of firearms with highly similar bullet rifling marks, while attaining 100.0% accuracy for firearms exhibiting significant differences in rifling characteristics, demonstrating a clear superiority over other algorithms. This study paves the way for further research and development in the field of forensic firearm examination through AI-driven solutions, aiming to improve the efficiency and accuracy of firearm identification and investigation processes.