A novel training methodology for phase segmentation of steel microstructures using a deep learning algorithm

Abstract

In this work, an efficient training methodology for the segmentation of ferrite – pearlite microstructures using Convolutional Neural Network (CNN) UNET machine learning architecture (a semantic classifier) is presented. However, CNN driven image segmentation requires very large number of labelled training microstructure, which is generally not available. A novel method is proposed for circumventing the above problem. First polycrystalline templates are created by simulating a 3D nucleation and growth model which follow Avrami kinetics. Subsequently, cropped images of pearlite and ferrite (from real microstructures) are randomly positioned on the individual grains of the polycrystalline templates, producing synthetic microstructures of varying fractions and varying scale of the two constituents. A few thousand synthetic microstructures were created using a small number of cropped images. The UNET trained on the synthetic training set when tested on real ferrite-pearlite microstructures gave an accuracy of about 98%, which substantiates the robustness of above technique. This paper introduces a universal training methodology for practical and efficient multiphase microstructure quantification based on supervised learning.