A Dual-Branch Based Pipeline Welding Defect Detection Method

This article introduces an innovative dual branch detection method aimed at improving the fine-grained detection capability of pipeline welding defects. The widespread application of welding technology in industrial production has accelerated the development process of pipeline manufacturing. In the process of promoting high-quality intelligent manufacturing, accurate detection of welding defects has become one of the key links. Due to the deep professional knowledge required for welding defect detection, existing methods have shown certain limitations in fine-grained defect detection. To overcome these shortcomings, this study proposes a diversified detection branch system that combines anchor-based and anchor-free complementary regression paradigms to obtain predictive information from multiple perspectives. Firstly, we constructed a backbone network feature encoder to obtain multi-level defect feature information. Then, we constructed a dual branch feature pyramid to perform divide and conquer prediction. Next, we will implement dual branch prediction using anchor based and anchor free approaches. In the experiment, a rich database containing multiple types of welding defects was used for model training and system testing, and multiple optimization indicators were comprehensively evaluated.