Supervised pearlitic–ferritic steel microstructure segmentation by U-Net convolutional neural network

Abstract The aim of this work is to develop an automated procedure based on machine learning capabilities for the identification pearlite islands within two-phase pearlitic–ferritic steel. input parameters custom implementation a braided neural network are provided as data set scanning electron microscopy images metallographic specimens. procedures related processing and optimization affecting final architecture effectiveness stage examined. objective find best solution problem ferritic–pearlitic microstructure segmentation, allowing further during, e.g., 3D reconstruction from serial sectioning. examines various quality different U-Net architectures one that can identify with highest precision. Two types acquired secondary (SE) backscattered diffraction (EBSD) detectors used during investigation. revealed developed approach offers improvements in investigations by removing requirement expert knowledge interpretation image prior characterization. It has also been proven artificial networks deep process using extensible nonlinear tools microstructure, while overtraining level remains low. Convolutional do not require manual feature extraction able automatically appropriate search functions recognize structure areas training without human intervention. was shown recognizes analyzed steel satisfactory precision 79% EBSD 87% SE images.