Exploring the potential of neural networks for firearms classification on bullet images

The two main points of attention raised in 2016 by the PCAST report for the firearm examination, are the importance of exploiting automation and the need to reduce subjectivity in the method. In response to these criticisms, we seek, in this research, to exploit deep learning algorithms that both reduce the subjectivity and enable the process of large amounts of data. More specifically, Convolutional Neural Networks (CNN), capable of processing raw images, have been implemented to reduce human input in the classification of firearms of the same model. To do this, we used three datasets, each containing firearms of the same model, to train the different models. Each model was then evaluated on its ability to correctly classify bullet images. The results, which varied greatly for the same neural network depending on the dataset used, highlight the importance of the structure of the dataset on the performance of a model. The overall prediction accuracies ranging from 24.25% (on 200 classes) for certain classifier to 100.00% (on four classes) for different models, show us the fragility of neural networks, mostly on the dataset employed. Finally, the use of neural networks remains still promising for the classification of firearms of the same model, but above all highlights the importance of future research in this field.