In-depth analysis of automated baggage inspection using simulated X-ray images of 3D models

X-ray baggage inspection ensures transport and border security, as it prevents hazardous objects from entering secure areas. Currently, deep learning is the state-of-the-art approach for automated threat object detection and classification. These networks require extensive training data; however, the number of publicly available datasets of X-ray images is limited. To overcome this, we propose an image generation pipeline that generates new data by superimposing simulated X-ray images of 3D models onto real baggage X-rays. This approach allows researchers to train deep neural networks without requiring additional imaging or manual labeling. The effectiveness and reliability of our image simulation pipeline are demonstrated, integrating advanced techniques such as distortion with diffusion models. We conducted hundreds of YOLOv5 trainings with a combination of real images from the SIXray dataset and simulated X-rays containing wrenches and handguns. Testing was performed exclusively on unaltered real images. Training exclusively with 16,000 simulated images of grayscale wrenches resulted in an \(AP_{0.5}\) $A{P}_{0.5}$ of 72.7%. For the handguns, using only 50 real images yielded an \(AP_{0.5}\) $A{P}_{0.5}$ of 78.8%; however, by adding 16,000 simulated X-rays to these real images, the \(AP_{0.5}\) $A{P}_{0.5}$ increased to 91.6%. Our results prove that using simulated images of threat objects can improve the performance of object detection models. As modern object detectors process images in real-time, they establish themselves as a feasible approach for aiding inspectors and even fully automating baggage inspection. Our novel superimposition and colorization techniques are not only relevant to security but can also be employed in other areas of X-ray imaging.