Centerline-Guided Weld Annotation for PCCP: Methods and Application Performance

Semantic segmentation is a pivotal computer vision task, yet pixel-level annotation is labor-intensive and costly. For ribbon-like weld seams with ambiguous boundaries, polygon-based annotation suffers severe inter-annotator variability, an issue that has received limited attention in existing model-centric research. This study proposes a centerline-based annotation strategy for Prestressed Concrete Cylinder Pipe (PCCP) weld seams to improve efficiency, consistency, and model performance. Ground truth masks are generated via the Image-Guided Centerline-to-Mask Generation algorithm using slope-adaptive kernels to convert centerlines into pixel-level masks. During training and inference, multi-directional convolutional response maps are concatenated with RGB images as six-channel inputs and evaluated on UNet, R2UNet, AttUNet, UNet++, and R2AUNet. Centerline annotation reduces interactive clicks by 63.5% and the average per-image annotation time by approximately 68%, while improving inter-annotator consistency (Cohen’s kappa: 0.913 to 0.967). Across all five architectures, mean intersection over union improves by 8.68%, with residual-recurrent networks (R2UNet and R2AUNet) gaining up to 17.32% and 13.98%, respectively. In the evaluated setting, the strategy improves annotation efficiency, inter-annotator consistency, and segmentation accuracy; its engineering feasibility was preliminarily assessed through a three-month deployment on a PCCP steel-cylinder leak detection line, during which the missed-detection rate was 0.90% (95% CI: 0.19–2.62%).