A New Welding Seam Recognition Methodology Based on Deep Learning Model MRCNN

Abstract

Welding is a common procedure in industrial production processing. For offering accurate machining trajectories of industrial robot, stereo vision system is usually used to compensate the position error of welding seam caused by local overheat or laying position deviation. At present, the welding stereo vision system is normally used to get position of solder joints by scanning the whole profile of work piece, which helps compute cloud point position and then transform it mathematically from camera coordinate system to robot base coordinate system, thus to guide the robot and welding system. But in vision guided system, the laser sensor can only move along the straight line of curves with curvature slowly changing. To solve it, this work provides a welding seam recognition model based on Mask R-CNN network, which uses transfer learning to identify welding seam in the image and segmentation of instances, and the method can recognize the complex welding seam effectively so as to obtain the accurate position of solder joint.