Machine-Learning-Based Evaluation of Ferrite Grain Size from Microstructural Images

A deep-learning approach is proposed for automatic measurement of ferrite grain size in ferrite–pearlite steel by the intercept method specified in JIS G 0551. Unlike conventional image-analysis methods that estimate grain size from segmented areas, the proposed method directly classifies each pixel on a one-pixel-wide horizontal test line into six classes, reproducing the manual intercept procedure. SAM2-UNet, which combines the pre-trained encoder of SAM2 with a U-Net decoder, is used as the segmentation model. By explicitly training the model to recognize etch pits, polishing scratches, and inclusions, confusion with grain boundaries is suppressed. A composite loss function of weighted cross-entropy and Dice loss, together with an edge-weight mask, is introduced. For 27 specimens of case-hardened steel, the proposed method achieved a correlation coefficient of r = 0.946 with manual measurement and a mean absolute percentage error of 9.39%.