An edge-based online mandrel surface defect inspection system under strong industrial interference
Аннотация
High-speed online inspection of mandrel surfaces is pivotal to ensuring the quality of seamless steel pipes, yet it is hindered by production line speeds up to 3 m/s and severe interference from pseudo-defects like water stains and graphite residues. To tackle these challenges, we propose a fully automated edge-computing-based 360° visual inspection system for mandrel monitoring on high-throughput production lines. Integrating multi-camera image acquisition with distributed edge nodes, the system delivers deterministic and low-latency responses. At its core lies the Anti-Interference Real-time Segmentation Network (AIRSegNet), tailored to robustly segment submillimeter-scale defects from complex backgrounds on edge devices. On our custom-built Mandrel Surface Defect under Strong Industrial Interference (MSD-SII) dataset, AIRSegNet achieves an inference speed of 371 frames per second on Jetson edge devices, meeting real-time inspection requirements, while attaining a mean Intersection over Union of 0.57 and a false-positive rate of 0.0011. These results demonstrate that the proposed system enables online monitoring of mandrel surface defects under complex industrial interference conditions. Code and dataset: https://github.com/YitengLi-0905/A-Anti-Interference-Real-time-Segmentation-Network and https://zenodo.org/records/22605606 .