A computer vision-based approach for identification of non-metallic inclusions in the steel industry products

Abstract

Identification of microstructures is the core of materials engineering. Artificial intelligence's application in materials engineering has recently shown the possibility of realizing complicated tasks. Identifying elemental distribution in microstructure requires experimentation or computationally intensive modeling techniques. The current work focuses on the question, can artificial intelligence predict elemental distribution in a microstructure? The case study was selected from the steel industry. Making steel will cause different inclusions; identifying them is essential for qualifying the steel for applications. The current study develops a unique computer vision-based architecture by integrating Swin Transformer and U-Net architecture to identify the inclusions. The developed model can predict the type of inclusion in the steel by generating the elemental distribution images. The model is compared with the possible available architectures in the literature. The new model shows the lowest mean absolute error of 0.0529, root mean squared error of 0.0902, mean squared error of 0.0081, and the highest structural similarity (SSim) value of 0.68965 and an intersection over union (IoU) of 1 when images are binarised.