The geometric characteristic parameters of the weld pool can reflect the welding quality in the laser welding process. Laser welding vision system can obtain the geometric parameters of molten pool in real time, which is very important for welding quality monitoring and welding automation. In this paper, the visual system of high-power fiber laser deep penetration welding process detection is improved. The image processing algorithm is systematically analyzed, and the edge of molten pool image is effectively extracted to avoid large interference, and realize accurate on-line measurement of weld pool geometric characteristic parameters. A method of closed convex active contour model based on prior information of molten pool is proposed. It can not only realize the accurate on-line measurement of the geometric characteristic parameters of molten pool, but also has better flexibility and adaptability. Compared with the balloon model algorithm, the measurement error of the closed convex active contour model is stable, the maximum error is less than 13.15%, and the balloon model measurement error is higher than 132.92%, and the error is very large. This method has a good effect of removing interference, and the algorithm is fast and can accurately monitor the geometric parameters of molten pool on-line.