Go to DBLP homepageAI-Based Generative Model for Computer-Based X-Ray Inspection Training System
Abstract: X-ray baggage inspection systems have proven to be indispensable tools for swiftly and efficiently examining passengers’ belongings and inspecting goods during import-export processes at locations, such as airports and seaports. The ability to visualize the contents of bags without opening them streamlines the inspection process, enabling rapid assessments. The reliance on human judgment for interpreting X-ray images requires systematic training program, given the disparities between the actual shapes of items and their representations in X-ray images. In addition, the surge in travel demand and logistics movements necessitates a substantial number of highly trained inspectors. Computer-based training (CBT) systems offer a convenient solution for training inspectors, but current X-ray inspection CBT systems have limitations in data diversity and provide only restricted information, constraining the effectiveness of training. To address this, here we propose a novel AI-based data augmentation scheme for X-ray inspection CBT systems. Specifically, this article focuses on the AI generative model component, which is composed of three parts. First, we employ a latent diffusion model (LDM) to generate high-quality, diverse illicit item X-ray images. Second, by incorporating a novel neural support optimization from frontal and side-view images, a tuned voxel grid is obtained, enabling the creation of 3-D images that offer varied perspectives of items. Finally, the generated images undergo transformations, such as metal material emphasis, organic material emphasis, negative images, and variations in brightness to enhance detection capabilities, making it easier to identify items that were challenging to detect in conventional pseudo-color images. Inspector training with real X-ray data followed by training with generated X-ray data resulted in an increase in inspection accuracy and decrease in inspection time, confirming the effectiveness of training using generated images.
External IDs:dblp:journals/tim/KimKRY25
Loading