Abstract: In order to meet the need of tracking underwater seam,a vision system was designed.In this system,a halogen tungsten lamp is used to light the groove in the front of arc and compound filters are selected.As a result,the comparatively clear welding seam images can be captured with the system.Meanwhile,an operator is proposed based on neighbourhood average value of edge image,and an improved genetic algorithm is used to calculate maximization of interclass variance for images,and the best threshold segmentation was obtained.The algorithm not only preserves the weak edge of image,but also reduces the running time.For welding seam image after segmentation,an algorithm based on properties of welding seam width was proposed which has overcome the disturbance of arc,splash and can recognize the seam and obtain the welding seam centre line accurately.The research laid the foundation for the automation of underwater welding.