Augmenting Data with GANs for Firearms Detection in Cargo X-Ray Images

We propose a framework and impact of applying Machine Learning-based generated imagery to augment data variations for firearm detection in cargo x-ray images. Deep learning-based approaches for object detection have rapidly become the state-of-art and crucial technology for non-intrusive inspection (NII) based on x-ray radiography. The technology is widely employed to reduce or replace tedious labor-intensive inspection to verify cargo content and intercept potential threats at border crossings, ports, and other critical infrastructure facilities. However, the need for variations in the threat cargo content makes accumulating training data for such a system an increasing development cost. Even though threat image projection (TIP) is widely employed to simplify the process into artificially projecting the known threat, a considerable amount of threat object appearances is still needed. To further reduce the cost, we explore the use of GenerativeAdversarial-Network (GAN) to aid dataset creation. GAN is a successful deep learning technique for generating photo-real imagery in many domains. We propose a three-stage training framework dedicated to firearm detection. First, GAN is trained to generate variations of X-ray firearm appearance from binary masks for better image quality compared to the commonly used random noise. Second, the detection training dataset is created in combinations of generated images and actual firearms using TIP. Finally, the dataset is used to train RetinaNet for the detection. Our evaluations reveal that GAN can reduce the training cost in increase detection performance as using the combination of the real and generated firearms increase performance for unseen firearms detection.