Gunshot caliber classification using convolutional neural networks

In this work, the acoustic classification of raw high and low caliber gunshot sounds using deep learning is investigated. Of particular interest is employing a Convolutional Neural Networks (CNN) prediction model. The raw gunshot sounds are preprocessed to separate the gunshot muzzle blast from other ambient sounds using Blind Source Separation(BSS) implemented using Short-time Fourier Transform (STFT) and Non-Negative Matrix Decomposition (NMF) to the magnitude’s spectrogram. The proposed model has an accuracy of 97%, precision, recall, and F1-score of 97%,99%, and 98%, respectively.