A lightweight pipeline weld radiographs defect detection detector based on Inverted Bottleneck

Pipeline weld defect detection is essential for the safety of oil and gas transportation, but visual radiograph inspection is time-consuming and labor-intensive. Intelligent detection often requires large, complex models with limited field applicability, and balancing the number of defect types, accuracy and lightweight is challenging. To solve the above problems, we propose a lightweight and efficient weld radiographs defect detector based on YOLO V8. By designing a C2f_DP module with depthwise convolution-expanded convolution layer-intermediate depthwise convolution-projected convolution layer structure to extract high-dimensional spatial information features while reducing parameters. And a Lightweight Cross-Scale Feature Fusion Module (LCCFM) structure with mixed local channel attention is proposed to redesign the neck network, which compresses the network size while adding feature transmission paths to efficiently integrate detailed features and contextual information. Furthermore a Focus and Inner-Minimum Point Distance based Intersection over Union(FI-MPDIoU) is proposed to enhance the generalisation ability of the algorithm and attenuate the negative impact of low-quality samples. A dataset containing seven types of defects is established to verify the effectiveness of the proposed method, the results show that the proposed method achieves a mAP of 92.7%, which is improved by 2.9%, and the parameters are decreased by 50.8%, GFLOPs are decreased by 35.8%, model volume is decreased by 47.8%, effectively improving the model performance and demonstrating its ability to be applied to the detection of welding defects in long-distance oil and gas pipelines.