<p>Turbine blades in aero-engines contain intricate internal cooling structures embedded within dense superalloy materials, making reliable non-destructive evaluation (NDE) particularly challenging. Industrial computed tomography (CT) is widely used for internal inspection; however, metal-induced artifacts and resolution degradation often disrupt structural continuity, especially in sub-millimeter regions, limiting defect visibility and measurement accuracy. To address this challenge, this study proposes a lightweight, two-stage structure-aware CT image restoration framework tailored for industrial inspection. In the first stage, a graph neural network (GNN)-based artifact suppression module is developed to capture non-local structural dependencies via dilated convolutions and residual refinement, thereby enabling effective mitigation of streaks, rings, and beam-hardening artifacts. In the second stage, a pre-trained Enhanced Super-Resolution Generative Adversarial Network (ESRGAN) is applied as a fixed post-processing module to enhance high-frequency details and local contrast without domain-specific retraining. Experimental results on real turbine blade CT data demonstrate that the proposed framework outperforms conventional filtering methods and convolutional baselines, including CNNs and U-Net, in terms of both PSNR and SSIM. The results further indicate that structure-aware restoration is essential for maintaining image fidelity under severe artifact conditions. By improving structural clarity, the proposed method may support downstream tasks such as defect localization and dimensional verification in industrial inspection workflows.</p>