Artifact Removal and Defect Identification in High-Resolution CT Images of CFRP Electrical Composites Using an Improved UNet Architecture

Carbon fiber reinforced polymer (CFRP) composites are widely utilized due to their exceptional electrical conductivity and electromagnetic shielding properties; however, defects generated during their application can compromise material performance. Micro X-ray computed tomography (μ-CT) serves as an effective technique for characterizing such defects, yet artifacts in reconstructed images are frequently misidentified as genuine defects, thereby affecting defect recognition accuracy. To mitigate artifact interference and effectively identify defects, this study proposes an improved UNet deep learning network. The approach employs a sliding window strategy to process high-resolution images, overcoming memory limitations, and integrates attention gating mechanisms with residual connections to enhance feature extraction capabilities, thereby constructing a defect identification model suitable for high-resolution CFRP images. The results of image-based defect identification demonstrate that this method effectively suppresses artifact interference, significantly reducing misjudgment of defects caused by artifacts.