To detect the defects on a smooth metal surface, a vision system in the condition of parallel lighting was built, and a lighted LED array was put as background of the target artifact. First of all, images were captured from different angles by three cameras, because the gray feature of the target surface and the defects between the brightfield image and the darkfield image is on the contrary. Then image processing algorithms were carried out to pick up the center coordinates of every LED point, which can be used as feature points in the DLT process. Finally differential fusion between brightfield and darkfield was implemented. Experiments show that differential fusion can either eliminate background interference information effectively or improve the contrast between the target surface and defects obviously.