An Ultra-sparse View CT Imaging Method Based on X-ray2CTNet

The prediction of the internal ballistic performance of solid rocket motors (SRMs) mainly depends on the observation of burning surface regression. The existing techniques have not yet achieved real-time, accurate and intuitive observations of the grain structural dynamic changes under hot conditions. Computed tomography (CT) imaging based on X-ray can be directly used to analyze the burning surface of SRM. However, neither the conventional CT imaging mode nor the current reconstruction algorithms, including full-view and sparse-view reconstruction methods, can be used in ignition tests. It is necessary to break through the limitation of sparse sampling and research ultra-sparse view reconstruction. Although computer vision technology has shown that 3D shapes can be estimated from very few 2D RGB images via deep learning, it remains challenging to reconstruct volumes from two 2D X-ray images. To tackle this issue here, we propose a 3D reconstruction network based on ultra-sparse projection views, namely X-ray2CTNet, which takes the 2D projections of any two orthogonal views as inputs to implement cross-dimensional inverse mapping from 2D (X-rays) to 3D (CT). In addition, to solve the insufficient dataset problem, training sets are constructed by simulating different models of grain regression for a certain type of propellant and utilizing a CT simulation platform. We pour fake grains that are the same as those of the simulation models to acquire the actual projections of two views for reconstructing the 3D volume. The obtained results prove the feasibility of our method. The proposed method provides a possibility for dynamic monitoring of burning surface regressions for SRMs in ground ignition tests.