Rifle Detection and Performance Evaluation Using Deep Learning Frameworks

Abstract

Deep learning models being used to improve human life has been an ongoing domain of research. Violence, especially with the proliferation of arms, has been on the increase worldwide. Many tragedies have occurred right across the globe, leading to people losing their lives as a result of being shot at with guns. This research sought to use deep learning frameworks to detect rifles in images and assess their performance based on the metrics of accuracy and F1 score. The study used a combination of images from Google open images and other sources to form a dataset of 2105 images; 1857 of those was used to train YOLOv3 and RetinaNet models to detect rifles, using Darknet-53 and ResNet50 respectively as the backbone networks. The models were evaluated after training using a test dataset containing 248 images, both the training and evaluation of the models were carried out using scripts written in Python. The results obtained showed that YOLOv3 had better output in terms of accuracy, precision, recall, and, consequently, the F1 scored better than RetinaNet