Synthetic data has emerged as a critical piece of the machine-learning based approach to X-ray system design and automatic threat recognition development. Physics-based synthetic data integrates virtual models with a physics-based simulation engine, thereby granting users the capability to produce synthetic measurements based on arbitrary, user-specified input objects and materials. Such inputs can range from geometric phantoms that assist in system design to new threat materials and configurations that expedite ATR training in response to emerging threats. We introduce enhancements to the QSimRT virtual model generation pipeline. This incorporates the rapid creation of virtual models representing passenger luggage, stochastically generated electronics, and user-specified model variability for extensive ensemble production. We have employed these models in training ATRs within the aviation security domain. This presentation will discuss the model generation process, emphasize its pivotal features, and share preliminary results derived from the application of these models in ATR training.
|