Part-based model for visual detection and localization of gas tungsten arc weld pool

Observing the weld pool and measuring its geometrical parameters are important issues for automated and robotic welding, wherein the visual detection and localization of weld pool are critical steps. Previous methods of visual measurement of weld pool usually assume that the weld pool exists in a predefined area, and its contour should be a specific geometric shape and size. Furthermore, previous methods were only suited for the pool images with complete boundary information and with small disturbing/noise edge. When part of the pool boundary is seriously spoiled (for example, by reflection) or confused by pileup area, it is very difficult if not impossible to conduct the geometric measurement of weld pool. In this paper, we propose a robust visual detection and localization method for the pool of gas tungsten arc weld based on part-based modeling and recognition of objects. It provides an elegant framework for representing the outline of a weld pool and is especially efficient for weld pool detection and localization in cluttered scenes, when it is partially occluded or when similar-looking pileup area can act as additional distracters. Experiments on real images verified the proposed method.