Weapon Categorization Using Artificial Intelligence

Wong, Ka Shing (2022) Weapon Categorization Using Artificial Intelligence. Final Year Project (Bachelor), Tunku Abdul Rahman University College.

Text Wong Ka Shing.pdf Restricted to Registered users only Download (4MB)

Abstract

Easy acquisition of handheld weapons has undeniably increased the crime rate. The lack of consistency and efficiency with manual detection of weapon possession calls for the need to develop an automatic weapon detection system aided with artificial intelligence. The objective of this research is to develop a deep learning model to detect and classify handheld weapons. Convolutional Neural Networks (CNN) is the favorable method to be applied to classification systems since it is a ‘learnable’ network. CNN can extract features, especially spatial data and form feature maps to identify the input efficiently. CNN is a good model in classification, owing to its weight sharing feature. The weight sharing feature can reduce the number of trainable parameters in the network, which in turn avoid overfitting and improve generalization. The CNN architectures employed in this work are InceptionResNetV2, InceptionV3, Renet50, Resnet101, SqueezeNet and VGG16. The weapon classification can classify 4 weapon categories, namely the rifle, pistol, knife and non-weapon. The performance of the CNN models is enhanced via the manipulation of various hyperparameters. The hyperparameters selected for the CNN models are 0.5 (non-augmented model) /0.8 (augmented model) dropout rate, 0.001 learning rate, 20 (non-augmented model) /50 (augmented model) epochs, SGD optimizer and Categorical Cross-entropy loss function. Performance evaluation of all six models in weapon categorization without transfer-learning suggests that VGG-16 provides the best performance (86% average F1 score and 95% average accuracy). Meanwhile, with transfer learning, InceptionV3 provides the best performance with a short training time (97% average F1 score and 99% average accuracy). Besides, Contrast Limited Adaptive Histogram Equalization (CLAHE) is performed as a pre-processing step to manipulate the contrast of the images. The result shows that CLAHE generates a comparable average score and F1 scores with the model without CLAHE applied, and is not a compulsory step. An additional non-weapon category is added to introduce incremental training to the CNN model. The results show that minor catastrophic forgetting occurs for the augmented model, nonetheless, it is not drastically affecting the overall classification capacity. It is established that the as-developed CNN model is reasonably robust against the challenge of increment learning, and is capable of classifying the dataset into 4 different categories – rifle, knife, pistol and non-weapon