Accurate cladding segmentation is essential for quantitative quality assessment of stainless-steel/carbon-steel clad wire rods used in bridge cables, yet remains challenging because of weak core–cladding contrast, narrow interfacial transition zones, local cladding-thickness fluctuations, and limited repeatability of manual inspection. This study proposes an improved U-Net framework that integrates residual feature extraction, multi-scale contextual perception, and attention-guided feature refinement for robust cladding identification. A cross-sectional image dataset comprising 18,566 samples was constructed through standardized specimen preparation, chemical color development, image acquisition, pixel-level annotation, and data augmentation. In the proposed model, the original U-Net encoder is replaced with ResNet50 to enhance deep semantic representation, while atrous spatial pyramid pooling and a convolutional block attention module are embedded into the feature-fusion stage to improve boundary discrimination and thin-cladding recognition. On the test set, the model achieved a mean pixel accuracy of 97.29%, cladding intersection over union of 88.82%, and mean intersection over union of 93.72%, outperforming the baseline U-Net by 1.38, 9.19, and 5.17 percentage points, respectively. Ablation and comparative experiments further demonstrate improved boundary continuity, local-detail preservation, and segmentation stability compared with representative CNN-based segmentation models. These findings suggest that the proposed framework provides a practical and reliable vision-based approach for cladding-thickness measurement, eccentricity evaluation, uniformity assessment, and batch quality inspection of clad wire rods.
Zeng et al. (Tue,) studied this question.