Combining Multi-Scale U-Net With Transformer for Welding Defect Detection of Oil/Gas Pipeline

Accurate welding defect detection (WDD) of Oil/Gas pipelines (OGP) is an active and challenging task in the reliability engineering of OGPs. To solve the problems that U-Net cannot effectively extract multi-scale global context details of the image by simple skip connection, and small defects cannot be accurately detected, a WDD method of OGP by combining residual-dilated-Inception U-Net (RDIU-Net) and Transformer (RDIUTrans) is proposed. In the model, RDIU-Net is used to extract the multi-scale local features, and Transformer is utilized to model multi-scale global contextual relationships and spatial dependency. Compared with U-Net and its variants, RDIUTrans can extract the global feature and local detail features for WDD. The results on the welding defect image dataset show that RDIUTrans is effective for WDIS with the segmentation accuracy of 95.34%. It is suitable for WDD scenes with various welding defects.