The roof of a car is a streamlined shape, and its weld has a high and low drop, so the brightness of the images taken at each position is different. The randomly generated air hole defects also have different characteristics, so it is difficult to formulate rules for detecting weld air hole defects in this situation.Using the Segmentation technology of Solomon SolVision AI image platform, the weld air hole defect images of various brightness are annotated to train the AI model, which can detect air hole defects anywhere on the weld regardless of the image brightness.