Electron tomography based on scanning transmission electron microscopy (STEM) is used to analyze 3D structures of metal nanoparticles on the atomic scale. However, in the case of supported metal nanoparticle catalysts, the supporting material may interfere with the 3D reconstruction of metal nanoparticles. In this study, a deep learning-based image inpainting method is applied to high-angle annular dark field (HAADF)-STEM images of a supported metal nanoparticle to predict and remove the background image of the support. The inpainting method can separate an 11nm Pd nanoparticle from the θ-Al2O3 support in HAADF-STEM images of the θ-Al2O3-supported Pd catalyst. 3D reconstruction of the extracted images of the Pd nanoparticle reveals that the Pd nanoparticle adopts a deformed structure of the cuboctahedron model particle, resulting in high index surfaces, which account for the high catalytic activity for methane combustion. Using the xyz coordinate of each Pd atom, the local Pd-Pd bond distance and its variance in a real supported Pd nanoparticle are visualized, showing large strain and disorder at the Pd-Al2O3 interface. The results demonstrate that 3D atomic-scale analysis enables atomic structure-based understanding and design of supported metal catalysts.