A Comparison of Six Convolutional Neural Networks for Weapon Categorization
Ka Shing Wong, Lih Poh Lin · 2022
This research aimed to employ artificial intelligence to classify handheld weapons. Convolutional Neural Networks (CNN) was the favourable method in which architectures including Inception-ResNet-V2, InceptionV3, ResNet50, ResNet101, SqueezeNet and VGG-16 have been applied. The performance of the CNN models was enhanced via the manipulation of various hyperparameters such as dropout rate, learning rate, number of epochs, loss function and optimizer. The performance matrices of all models in weapon categorization were promising, in which their performance in terms of precision, recall, specificity, F1 score and accuracy were all above 95%, except for SqueezeNet which achieved performance ranging from 75% to 90%. InceptionV3 appeared to have an upper hand with an accuracy of 99%. The CNN model was also robust against the challenge of increment learning in which it was able to classify weapons despite the addition of non-weapon classes and the occurrence of catastrophic forgetting. Given its promising performance, the application of CNN in weapon classification is much anticipated.