Automatic vision-based firearm recognition using deep learning approaches: a comparative analysis

Abstract

Firearms held up in an environment not designated for legal firearm activity is an indicator of a dangerous situation. This research proposes a vision-based firearm detection model which is able to assist in such a scenario in public spaces such as malls, airports, and residential areas. The aim is to evaluate the performance of various deep learning architectures based on their execution of detecting visible firearms in still images. The approach involves using publicly available datasets for deep learning detectors to learn from by extracting features related to firearms. Detection results are compared with one another based on the F1-score, a function of precision and recall measures for detector evaluation. The deep learning architectures to be evaluated are a mixture of existing state-of-the-art single-stage and two-stage object detectors typically used in deep learning. These include the Single Shot Multi-box Detector (SSD), You Only Look Once (YOLO), Region-based Fully Convolutional Network (RFCN), and Faster Region-Convolutional Neural Network(Faster R-CNN). The results obtained by the detectors describe how accurately each one performs the classification and localisation tasks in comparison with one another. Further-more, inferencing speed of each detector after being fully trained is also compared