An appearance inspection method for resistance spot welding based on semantic segmentation

Abstract

Resistance spot welding (RSW) plays an important role in manufacturing. The quality of the welding can be efficiently assessed by its appearance. Image segmentation is an important part of the RSW appearance inspection. However, the classical image segmentation algorithms cannot work very well because of the various RSW appearances. In this study, a novel inspection method is proposed based on semantic segmentation. We choose MobileNetV2 as the backbone for the semantic segmentation. After modification and optimization of the network, our model achieves an accuracy of 89% mean intersection-of-union (mIOU), which is averagely 30% higher than the classical image segmentation algorithms. A classifier further evaluates the quality of the RSW according to some geometric features of the segmented regions, and the classification accuracy is improved by 0.79%. This research is of great importance for the high accuracy quality control of the massive production to reduce the producing cost and improve the efficiency of the RSW pipeline.