DeepKP: A Robust and Accurate Framework for Weld Seam Keypoint Extraction in Welding Robots

Abstract

To meet the demand for seam tracking of welding robots, a deep learning-based framework called DeepKP is proposed in this paper, which aims to precisely extract weld seam keypoints under multiple arc light interference. DeepKP comprises a keypoint extraction model named WeldExt and a denoising model named WeldDenoise. WeldExt is proposed to identify weld seam types, obtain prior box region, and then extract keypoints in the region. WeldExt addresses the problem that most recent extraction models cannot directly obtain the keypoints. WeldDenoise is proposed for denoising weld seam images affected by multiple arc light interference, which overcomes the limitation that most recent denoising models must use paired datasets for training. Experiment results show that the average locating error of weld seam keypoint in WeldExt is 1.75 pixels and the average seam tracking error of DeepKp is 0.336 mm. Therefore, DeepKP performs excellently in extracting weld seam keypoints under challenging arc light conditions and improves the quality of seam tracking.