Employing data generation for visual weapon identification using Convolutional Neural Networks

Abstract

The use of weapons nowadays is becoming a leading cause of severe crimes in our society, which reluctantly results in dreadful consequences. The weapons used typically varies from knife, iron-rod, dagger, sabre to firearms like guns and bombs. Due to the unavailability of any proactive mechanism for avoiding heinous crimes using such weapons, an active surveillance performing real time weapon identification is proposed here, as a boon to societal security requirement. As part of it, this paper presents a novel approach based on Convolutional Neural Network (CNN) for identifying visual weapons. This proposed CNN model is initialized with the pre-trained Visual Geometry Group-16 (VGG-16) network weights. These weights are further fine-tuned by training this CNN model with comprehensive weapons (knives and handguns only) and non-weapon images. Weapon category images correspond to further classified classes of “isolated” and “handheld” weapons. However, weapon identification is challenging because of unavailability of diverse databases containing images with variations in shape, texture, scale, occlusion of weapon, etc. This paper reduces this limitation by presenting an algorithm for generating new images and other algorithm for preprocessing the images for quality enhancement. The accuracy achieved is 98.07% with original isolated images and 98.36% with its preprocessed images, while 98.42% with original handheld images and 98.80% with its preprocessed images. The preprocessed algorithm’s applicability is confirmed by the higher accuracy achieved by this model using preprocessed images. The accuracy achieved is on an average of ~7% higher than those achieved by other researchers with similar work. The improved result of weapon identification in terms of accuracy proves the appropriateness of the proposed research in being used commercially.