Deep learning in forensic shotgun pattern interpretation-A proof-of-concept study

Abstract Little is known about the potential of artificial intelligence in forensic shotgun pattern interpretation. As shooting distance is among the main factors behind shotgun patterning, this proof-of-concept study aimed to explore the potential of neural net architectures to correctly classify shotgun pattern images in terms of shooting distance. The study material comprised a total of 106 shotgun pattern images from two discrete shooting distances (n = 54 images from 10 m and n = 52 images from 17.5 m) recorded on blank white paper. The dataset was used to train, validate and test deep learning algorithms to correctly classify images in terms of shooting distance. The open source AIDeveloper software was used for the deep learning procedure. In this dataset, a TinyResNet-based algorithm reached the highest testing accuracy of 94%. Of the testing set, the algorithm classified all 10 m patterns correctly, and misclassified one 17.5 m pattern. On the basis of these preliminary data, it seems achievable to develop algorithms that would serve as a beneficial tool for forensic investigators when estimating shooting distances from shotgun patterns. In the future, studies with larger and more complex datasets are needed to develop robust and applicable algorithms for forensic shotgun pattern interpretation.