Deep learning-based 2D-3D sample pose estimation for X-ray 3DCT

Abstract

3D X-ray Computed Tomography (CT) is increasingly being used for non-destructive inspection of objects. Conventional CT inspection requires many projections, typically spanning 360 to reconstruct a 3D image of the object, which is then segmented and subsequently compared with the reference computer-aided design (CAD) model. Such an inspection flowchart, however, is a time inefficient procedure, not suitable for inline inspection. To overcome this problem, we directly compare the measured projections with simulated ones from the CAD model. To do so, the simulated projections need to be created with the same acquisition geometry as the measured ones. When an object is inserted on a scanning system, its orientation may vary with respect to the default CAD model orientation. For this reason, 2D/3D registration between the CAD model and the measured projections of the real object is necessary. In this paper, we present a deep learning based method to accurately estimate the 3D orientation of an object from one projection image.