| Summary: | Object detection of prohibited items in X-ray security inspection is challenging because of serious overlap, disorderly background, and high throughput. In the past few years, a variety of deep learning algorithms have been proposed and achieved satisfactory results. However, the performance of these algorithms relies heavily on the specified datasets. Moreover, establishing a large-scale X-ray image dataset by manually collecting and labeling images is prohibitively expensive and time consuming. In this paper, we propose a text-driven framework for synthesizing X-ray security inspection images based on Generative Adversarial Networks (GAN). First, a conditional GAN is developed to generate natural images of prohibited items from class labels. Second, an improved model based on a pix-to-pix GAN is implemented to convert natural images into X-ray images. Third, another HD pix-to-pixel GAN is responsible for producing high-resolution benign background images, which are subsequently fused with the generated images of prohibited items to create X-ray inspection images. Finally, the proposed method is evaluated using SOTA object detection algorithms, such as YOLO-v5, and achieving 4.6% promotion for <inline-formula> <tex-math notation="LaTeX">$mAP_{0.5}$ </tex-math></inline-formula> and 15.9% promotion for <inline-formula> <tex-math notation="LaTeX">$mAP_{0.5-0.95}$ </tex-math></inline-formula>. The experimental results demonstrate that our image synthesis framework can effectively augment the datasets of prohibited items and improve the detection performance of deep learning algorithms during X-ray security screening.
|
|---|