An image analysis system for coaxially viewed weld scenes

Abstract

We present a complete, working system for analyzing coaxially viewed robotic weld scenes. The analysis is cast in the form of a consistent labeling problem in which the objects are small image regions from a regular tesselation and the possible labels are base metal, electrode, gas cup, filler wire, weld bead, and weld pool. Local domain knowledge and measurements on the image function are used to produce an initial set of labeling probabilities. These are then adjusted by probabilistic relaxation using global domain knowledge to arrive at a final consistent labeling, which constitutes the image analysis. The primary goal of this analysis is a sufficiently accurate description of the size and shape of the weld pool to allow quality monitoring of the welding process. This represents a significant departure from the primary objectives of prior work in this area, most of which has focused on the seam tracking problem. Discussions with welding engineers indicate that they are anxious to acquire whatever information may be available. Thus, this system serves the secondary goal of providing some idea of what might be possible, given both relatively modest, and therefore affordable, resources and a real time performance requirement. In constructing and demonstrating a complete system, we provide useful insight into the engineering of such systems for practical applications, addressing attribute extraction, feature selection, and statistical region classification. A novel, efficient near-optimal feature selection algorithm which we callratchet search is also presented. Finally, we discuss how such a system, which is quite robust, could be embedded into robotic welding systems to provide important weld quality analysis for process control.