A unified framework based on semantic segmentation for extraction of weld seam profiles with typical joints

Abstract

Weld seam profile (WSP) extraction is the crucial step in laser vision-based robotic arc welding. Faced variable joints and profiles and different interference background, traditional machine learning methods generally produce unsatisfactory extraction results in accuracy and the generalization and anti-interference abilities. This paper presents a unified framework based on an improved U-Net for WSP extraction with various typical joints, focusing on upgrading the performances above. In this proposed network, the channel attention feature fusion block is firstly designed to learn the global contexts of WSPs from high and low-level feature maps. Secondly, the residual spatial pyramid pooling-fast block is proposed to distinguish the features of laser stripes and spatter, via independent chunked-area pooling. Thirdly, the multi-scale bottleneck attention module is developed to realize accurate prediction of boundary pixels of WSPs, through accumulating multi-scale features. Finally, the dice loss function is used for model training to address the problem of the unbalanced category, involving WSPs and the background. Experimental results show that the proposed network can effectively extract various WSPs with strong arc background, spatter, and noise, covering butt, fillet, and lap joints. Also, this improved U-Net is superior to four classical models FCN, SegNet, Deeplabv3+, U-Net, and to our previous traditional methods, in terms of accuracy, and the generalization and anti-interference abilities. This network provides a unified framework for extracting WSPs with various joints, regardless of positions, directions, and shapes of WSPs in images. It shows potential for automatic and intelligent arc welding.