Preliminary evaluation of gunshot residue pattern analysis using flash-pulse infrared thermography and multi-task deep learning

Gunshot residue pattern assessment can support forensic reconstruction, particularly for shooting distance evaluation, but commonly used approaches may be time-consuming, costly, or dependent on subjective visual interpretation, which can be especially difficult on challenging substrates. At the same time, no prior study, to our knowledge, has combined flash-pulse infrared thermography with multi-task deep learning for simultaneous prediction of multiple ballistic attributes from thermographic deposition patterns on textiles. In this proof-of-concept study, we investigated whether these patterns associated with firearm discharge residues and damage on textile targets can support complementary ballistic screening under controlled laboratory conditions. Using a dataset of 312 ballistic samples, we developed a hybrid multi-task neural network to classify weapon category, firearm model, and ammunition type, while also estimating shooting distance. The final model achieved mean accuracies of 94.3% for weapon category classification, 85.1% for firearm model classification, and 92.6% for ammunition type classification, together with a mean absolute error of 7.92 cm for shooting distance estimation. The proposed approach outperformed the evaluated traditional machine-learning baselines and single-task neural-network models within this experimental setting. These findings support the feasibility of combining flash-pulse infrared thermography and multi-task deep learning as a rapid, non-destructive, and objective complementary tool for ballistic screening and shooting distance assessment in controlled conditions. However, because the study was limited to a restricted set of firearms, ammunition types, textile conditions, and laboratory-controlled samples, the method should be regarded as a preliminary screening approach rather than a confirmatory forensic identification method. • Flash-pulse thermography enables rapid non-destructive GSR analysis on textiles. • Hybrid model achieves 94% weapon category and 85% weapon model accuracy. • Multi-task learning simultaneously predicts weapon, ammunition, and distance. • First multi-task framework for GSR-based ballistic attribute classification.