Multi-class semantic segmentation for identification of silicate island defects

Abstract In the automotive industry, it is necessary to identify the edge and center silicate island weld defects formed during Gas metal arc welding. These inspections of the weld are typically performed manually by visually inspecting the weld and identifying regions where the defect concentration is greater than a set threshold. Such a system is prone to errors and can be time-consuming. A novel deep-learning neural network is required to meet the industry’s demand for high-quality welded products. To achieve this, a deep learning U-Net model for multi-class semantic segmentation was designed. The model was trained with a dataset of less than a hundred images and can achieve over 98% accuracy.