Developing a Cloud-Based Weapon Detection System Using Transfer Learning and Generative Adversarial Networks

Public safety is a massive concern for governments and civilians alike. Measures taken to ensure public safety range from urban and safe city planning at the government level to self-defence programs for the public. Among these measures, baggage checking and surveillance systems are already in place. The motivation for this paper is to enhance these existing systems. In this paper, the authors propose an object detection system for baggage security that will significantly automate the surveillance process. The proposed system uses deep learning techniques, such as Transfer Learning (YOLO and Detectron2 pre-trained models achieved loss values of 0.01345 and 0.04498, respectively), Generative Adversarial Networks-GAN (for synthetic image creation with a batch size of up to 32), and cloud computing. It aims to identify threatening objects and isolate them during the X-ray scanning process in baggage security. A large dataset is generated using the GAN technique and data augmentation to train the model. Threats are then detected using popular models such as You Only Look Once (YOLO) and Detectron2. Additionally, this paper proposes a cloud-based system to facilitate seamless conversion between multiple object detection annotation formats, such as PascalVOC, YOLO, and Common Objects in Context (COCO). The cloud-based system uses client-side calculations and compression techniques to convert annotation data interchangeably between the mentioned formats, providing users this service free of charge.