Microstructure Segmentation and Quantification of Advanced Steels Combining EBSD and Deep Learning

Chunguang Shen, Chenchong Wang, Wei Xu*

Northeastern University, Shenyang, 110819, China

ABSTRACT: Recently, some material experts have started exploring the application of data driven-based approaches for microstructure segmentation and further quantitative metallography. However, no researcher has yet applied such approach to actual engineering steel with complicated microstructures limited by the lack of accurate ground truth for model training. To address above issues, an EBSD-assisted deep learning (DL) framework was proposed to classify complicated microstructures, wherein EBSD technique was introduced to accurately determine phase species for building high-quality dataset and U-Net network was applied to establish the correlation between input image and ground truth. In present work, DP steel and QP steel were selected as the target materials with complex microstructures. Two small datasets were built based on the SEM and EBSD experiments. The U-Net architecture was selected to recognize microstructures via segmentation of input image considering its good performance in some complex recognition tasks using only limited data. The segmentation results of DP steel were shown in Fig.1a. It can be clearly observed that although some martensite contains different morphological features, these phases can be also accurately classified using present model. The segmentation results of DP steel were shown in Fig.1b. With the guidance of EBSD technique, ferrite, RA and martensite in Q&P steel were classified using DL model based on reliable ground truth and acceptable predicted accuracy was obtained. Based on accurate DL model, an automatic quantitative method was proposed using image processing method based on Opencv toolbox. The deviation of the calculated volume fraction between EBSD and present method is less than 5%.

Figure 1. Segmentation results: (a) DP steel; (b) QP steel

Keywords: Microstructure quantification; EBSD; Deep learning