The use of underwater welding is on the increase as more and moreoffshore structures are being fabricated. In the need of exploring theresources in the deep sea, underwater welding is also used in that deepunder sea. It is strongly expected to weld automatically under that condition.Automatic seam tracking is an important part of weld automation. To trackseam automatically during underwater flux-cored arc welding, avision-based seam tracking system has been fulfilled. This seam trackingsystem concludes a vision sensing system, an image processing and seamrecognizing system, a fuzzy controller and an executive unit.Under the condition of arc welding, it is difficult to capture idea seamimage. By the analyzing of welding arc spectrum and based on theexperiments, a vision sensing system was built. In this system, ahalogen-tungsten lamp is used to lighting the groove in the front of the arcand a suitable filter is selected. Good seam image can be captured by thisvision sensing system.Due to the intensive arc light, randomly breaking of the bubble andgathering of the gas in the welding area, the seam images captured concludemany noises and they need to be processed before the seam is recognized.Noises in the seam can be filtered by a median filter. Many imagerecognition methods are based on the binary images. The binarization ofseam image are studied and an adaptive threshold adjusting method areworked out. A seam recognition principle based on the gray leveldistribution in spacial domain of binary image is showed in the paper. Whenthe seam image is in good condition, the recognition is accuracy by thisprinciple. While the seam image is in bad condition, the recognitionaccuracy decreases. The gray gradient at the edge of the seam is a great