Conditional generative adversarial network-based training image inpainting for laser vision seam tracking

Abstract

In the process of welding seam tracking based on laser vision, the location accuracy of the seam largely depends on the quality of the welding image training samples acquired in real time. However, the interference of a strong arc, spatter, and other noise in the welding process seriously contaminate the training images, which can cause tracking model drift and result in tracking failure. To enhance the robustness of the seam tracking model and improve the welding accuracy, a welding image inpainting method based on a conditional generative adversarial network (CGAN) is proposed. We constructed a welding image inpainting network and defined a loss function for the network training. Through training, the network learns an end-to-end mapping from a noisy welding image to the corresponding noise-free image. Then, to realize accurate automatic seam tracking, the optimized inpainting network was integrated into a tracker for training sample restoration, which improves the antinoise interference performance of the seam tracking system. The experimental results show that the proposed seam tracking method can stabilize the average welding error within 0.2 mm, which is superior to the existing methods. This demonstrate the effectiveness of the proposed method for improving the robustness and welding accuracy of the seam tracking system.