Abstract: Industrial cone-beam computed tomography (CBCT) is a core technology for nondestructive testing in fields such as aerospace, electronics, and automotive. However, scattering artifacts in its imaging can obscure critical defects such as microcracks and cold solder joints, thus significantly reducing detection accuracy. Most current scatter-correction methods rely solely on a single deep-learning technique, thus complicating alignment with actual physical effects and causing deviations in the Hounsfield unit (HU) value. This study proposes two scatter-correction methods: the deep ultrafast Boltzmann equation solver method for the projection domain and the deep artifact-compensation method for the image domain. Both methods use the UNet model instead of the conventional threshold segmentation to generate high-precision physical templates. The deep ultrafast Boltzmann equation solver method further combines the ultrafast Boltzmann equation to simulate photon physical effects, thereby achieving correction in the projection domain and completing correction in the image domain via a difference operation involving a physical template and low-pass filtering. This study evaluates both methods on datasets for copper-column crack detection and connector cold-solder joint detection. The deep artifact-compensation method requires no iterations and features simplified operations. Under physical-template calibration, it achieves an average root mean square error (RMSE) improvement of 9.5% compared with image-domain methods that rely solely on deep learning. However, its global smoothing results in the loss of high-frequency information. By contrast, the deep ultrafast Boltzmann equation solver method offers better detail preservation and correction accuracy, thus enabling a clearer restoration of fine structures. While ensuring structural consistency (with structural similarity index values of 0.96 and 0.93, respectively), the RMSE between the corrected images it generates and the reference images is only 11.38 HU and 32.64 HU, respectively, which corresponds to an average improvement of 23.5% over other methods. Ablation experiments confirmed the key role of the UNet model in improving template quality. This study provides an “accuracy–efficiency” option for industrial CBCT, thus contributing to enhanced reliability in nondestructive testing.