An image segmentation-based localization method for detecting weld seams

Abstract

In the process of assembling heavy vehicle rear axles, weld seams undergo meticulous quality checks. Research has predominantly focused on surface imperfections of weld seams, with minimal exploration into critical factors like the seam's exact location, configuration, and internal integrity. Precise seam localization is fundamental for a thorough assessment of weld quality and the advancement of welding techniques. In response, we have engineered a superior method for accurately localizing weld seams, even in the demanding conditions of industrial workshops. This method introduces an innovative attention-enhanced feature fusion network (AFFNet) that strengthens feature detection and accurately identifies weld seams down to the pixel level. To increase the accuracy of segmentation, we have developed a weighted dice loss function. Our methodology also includes a sliding window technique for identifying control points, which ensures smoother and more consistent localization of weld seams. The efficacy of our approach is validated through testing, showcasing outstanding performance with fewer parameters than other existing methods. The precision of our method is evidenced by the localization root mean squared error (RMSE) for both straight and curved weld seams, which are 1.72 and 2.09 pixels respectively, demonstrating the system's capability to precisely identify weld positions. This approach also meets the requirements for real-time inspection, marking a significant advancement in the field of weld seam analysis.