Phase Fraction Calculation in Multiphase Steel Microstructures Through Semantic Segmentation

The characterization of a material is defined by its transformation kinetics which is measured using the fraction of volume transformed. In common situations, a polycrystalline parent phase is transformed into various sub-phase constituent regions based on the desired material property. Characterization of such multiphase regions include accurate identification of phase and calculation of the phase fraction. To this end, we propose using semantic segmentation to identify and calculate the phase fraction in steel microstructures. We assign each pixel in an image to accomplish both phase identification and segmentation which we perform using enhanced ELU-Net coupled with improved combined loss function. Though scalable experiments we demonstrate more accurate phase fraction calculation via superior semantic segmentation by achieving an overall 8.6% performance increase over baseline with mean phase fraction error of 5.69%.