Closed-loop edge-server collaborative intelligence framework for corrosion assessment and 3D visualization of hydraulic steel structures

Current approaches to automated corrosion assessment of hydraulic steel structures (HSS) lack integrated deployment and visualization workflows and face challenges in fine-grained feature extraction, edge deployment, and annotation costs. This paper proposes a closed-loop edge-server collaborative framework for real-time corrosion assessment and 3D visualization. The lightweight YOLO26-CorroSeg network incorporates a dual-branch spatial-frequency module to capture subtle corrosion boundaries while maintaining hardware-acceleration compatibility, achieving 91.57% mIoU on the test set. A cascaded corrosion segmentation framework reduces reliance on pixel-level annotations of corrosion severity, while active sample selection uses spatial isomorphism between segmentation masks and Grad-CAM heatmaps to guide annotation. Edge-acquired data are transferred to a local server for offline reconstruction using an optimized 3D Gaussian Splatting pipeline. This study contributes an integrated approach connecting corrosion perception, annotation-efficient model development, and spatial visualization through edge-server task allocation, supporting automated infrastructure inspection and condition documentation.