Abstract: Seam tracking and weld penetration control are key parts of weld quality control. A three-light-path vision sensing system is used in the experiments to obtain the images of the top-front, top-back and back paths of the weld pool during the Al alloy GTAW welding and project them in the same picture at the same time. The image contains information of the seam, weld pool and back weld pool. The method of visual attention is adopted to find the small areas related to weld pool feature characters, and extract these characters from the image. The results show that, in real-time detection of the weld pool features during the welding process, the method of visual attention is more clarified and efficient than general methods as it focuses only on interested small areas.