Attention U-Net-based semantic segmentation for welding line detection

In industrial processes, quality assurance through methods such as visual inspection is essential for ensuring process stability. Traditional manual visual inspection is a time-consuming and costly endeavor. If the opportunity arises, replacing manual visual inspection with AI could lead to significant efficiency gains. However, simply judging the correctness or incorrectness of a process is often insufficient; quantitative attributes must also be associated with visual inspection. This paper proposes a solution for replacing manual visual inspection with AI specifically for welded joints. The aim is not only to detect the presence of weld joints but also to assess their geometric dimensions. Leveraging a proposed Attention U-Net architecture in combination with rule-based metrics, the proposed method offers a novel solution for identifying welding lines in images. By integrating semantic segmentation techniques, the method effectively distinguishes weld joint elements, while rule-based metrics facilitate the identification of critical cases requiring human intervention. Experimental results demonstrate the method's capability to automate a significant portion of inspection tasks, thereby reducing the reliance on manual labor and enhancing overall process efficiency and reliability.