Outer surface defect detection of steel pipes with 3D vision based on multi-line structured lights

Abstract

Abstract Surface defects pose a significant threat to the quality of hot rolled seamless steel pipes. While the efficacy of contemporary vision-based deep learning methodologies is undeniable, they encounter significant challenges in accurately identifying defects of substantial depth that compromise quality. Furthermore, these techniques often erroneously report numerous superficial defects. To overcome this obstacle, we have designed a novel visual detection system specifically for identifying surface defects on steel pipes. This system is inspired by laser triangulation and compensates for the absence of depth information in 2D images by leveraging the shape alterations of a multilinear structured light bar projected onto the steel pipe’s surface. Addressing the challenge of acquiring evenly distributed and difficult-to-obtain defect samples in real-world production processes, we have incorporated an unsupervised anomaly detection network, PatchCore, into the system. The proposed method achieves an area under the receiver operating characteristic curve of 99.84% and an F1 score of 0.9778 on a dataset collected and labeled at an industrial site. Furthermore, the online detection system has been successfully integrated into a hot rolled steel pipe production line, underscoring its practical applicability.