Evolution of VMamba models.

Welding defect detection of steel pipelines (SPWDD) is critical for ensuring the safe use of oil/gas pipelines. Due to the complex morphology of welding defect X-ray images (WDXI) in steel pipes and the/low contrast between the defects and the background, SPWDD is an important and challenging topic. To address these challenges, a Multi-Attention Cross-scanning VM-UNet (MACVM-UNet) for SPWDD is constructed. This model adopts the cross-Scanning Visual State Space Model (CVSS) to capture the local features and long-range dependencies, introduces channel attention skip connections (CASC) instead of the conventional skip connections to enhance the performance of globally and locally feature fusion, and employes the multi-scale Attention Feature Aggregation (MSAFA) module to fuse the multi-scale features. The combination of CVSS, CASC and MSAFA can effectively enhance the performance to extract the global-local features of small-sized and large-sized WDXIs. The experimental results on the WDXI dataset validate that the proposed MACVM-UNet outperforms the state-of-the-art models with superior SPWDD performance while maintaining low computational complexity. It achieves the mAcc 86.03%, mPre of 86.14% and mF1 of 84.6% with lower training time of 1.61 h. The proposed method provides an efficient and feasible solution for non-destructive SPWDD of oil/gas pipelines.