Pose estimation of multi-view X-ray radiography with variable imaging geometry for industrial assembly inspection

X-ray radiography, an essential technology for detecting the internals of objects, provides a vital position reference for terahertz travelling wave tube assembly, meeting the high-precision alignment requirement of internal devices. While existing deep learning-based pose estimation methods have made strides in achieving accurate results on X-ray images, they often require custom approaches for different system setups, making them unsuitable for assembly inspections where the imaging system geometry needs to be adjusted according to the assembly condition. This limitation underscores the need for a new approach. In response, we developed a novel CNN-based architecture called the Free-view Pose (FVPose) network, featuring system parameter embedding and attention mechanisms for multi-view feature integration, enabling flexible adaptation to varying imaging setups. Experimental evaluations demonstrate that FVPose outperforms state-of-the-art PnP-based networks in accuracy and efficiency while maintaining significantly faster processing speeds than traditional X-ray registration methods. These advancements position FVPose as a practical solution for real-time industrial assembly inspection applications.