Fig. 3. Light optical and scanning electron micrographs superimposed with lath-bainite predictions of different models and annotated regions showing the comparison between model prediction (red) and manual expert annotation (blue).

a, b Random initialized vanilla U-Net (model #1 and #6). c, d Random initialized U-Net VGG16 (model #3^v and #8^v). e, f Pre-trained U-Net VGG16 (model #3 and #8).