A multi-scale attention Siamese point cloud network for 3D similarity matching of firing pin impressions

The similarity matching of firing pin impressions on cartridge cases plays a critical role in forensic firearm identification. Traditional comparison methods, whether manual or geometry-based, often struggle to capture the subtle local variations and global structural patterns present in high-precision 3D impression data, leading to limited robustness and accuracy. In this paper, we propose a novel deep learning framework based on a Multi-Scale Attention Siamese Point Cloud Network to address these challenges. The proposed model integrates a PointMLP-based Siamese architecture with a multi-scale attention mechanism to jointly extract local geometric details and global contextual information from 3D point cloud representations of firing pin impressions. This design enables the network to effectively capture fine-grained differences between highly similar impressions, improving similarity discrimination capability. The framework is evaluated on a self-constructed dataset of 3D firing pin impressions, acquired through high-precision laser scanning from actual firearm discharges. The experimental results demonstrate that the proposed method outperforms traditional and existing learning-based approaches, achieving similarity matching accuracies of 98.91% on the training set and 99.30% on the test set. The approach offers a transferable solution for 3D similarity learning tasks, with potential applications in other 3D object comparison and forensic scenarios.