CNN-based Pose Estimation of Manufactured Objects During Inline X-ray Inspection

X-ray Computed Tomography (CT) is a nondestructive technique widely used for inspection of manufactured objects generated from a reference computer-aided-design (CAD) representation. In a conventional CT inspection framework, a volumetric reconstruction is computed from a large number of X-ray projections of the object. Afterwards, a surface is extracted, aligned and compared to the CAD model. For an accurate comparison, a high-resolution reconstruction is needed, requiring hundreds of projections, making this procedure not suitable for real-time inspection. In contrast to CT-based inspection, radiograph-based inspection only requires a few radiographs that then can be compared with simulated projections from the reference CAD model. For an effective comparison, however, an accurate 3D pose estimation of the object and consequent alignment between the measured object and the reference model are crucial. In this paper, we present an inline projection-based 3D pose estimation framework using convolutional neural networks (CNNs). Through realistic simulation experiments, we show that, with only two projections, estimation of the pose of the object is possible at the resolution of the acquisition system.