Radiographic inspection of Submerged Arc Welding using semantic segmentation

Submerged Arc Welding (SAW) is one of the most applied technology in manufacturing structural shape, larger diameter petroleum pipe, pressure vessels, storage tanks and machine components for all types of heavy industry. Digital radiography (DR) enables non-destructive evaluation of SAW welds. However, data provided by most industrial DR imaging systems suffers from low resolution and contrast, high noise background which makes manual inspection a challenging task. Unlike conventional methods aiming at detection of weld defects, We focus on pixel-wise extraction of both welds and defects out of the radiographic frames via semantic segmentation. Experiments suggest that our method yields quality pixel-level inferences: the size, form and positions of the welds and defects. This could be more meaningful for further evaluations where the relative positions and morphological feature of welds and defects are needed.