Abstract
Convolutional neural networks (CNNs) were used for the semantic segmentation of angular monocrystalline WC from NiCrBSi-WC optical microscopy images. This deep learning approach was able to emulate the laborious task of manual segmentation effectively, with a mean intersection over union (IOU) and a mean dice coefficient (DC) of 0.911 and 0.953, respectively, across the entire test dataset. From the model output, the carbide percent can be determined by dividing the area of positively labeled pixels by the total area of the image. Additionally, the mean free path can be quantified using the method described in ASTM STP 839, and by physically counting the black pixels (CPB) between the particles in the image. Comparing the models predictions to the ground truth, the carbide percent had an average difference of 1.2 area %, while the mean free path differed by 15.7 mu m for the ASTM method, and 24.8 mu m for the CPB method. The robustness of the model was tested on images containing both spherical eutectic WC and angular monocrystalline WC to determine whether the model was capable of accurately predicting the location of objects that were not part of the training dataset. The U-Net CNN was able to segment the spherical and angular WC with considerable accuracy. These results show that the application of computer vision models for microstructural characterization is not limited to complex imaging modalities, and can be applied to readily available methods such as optical microscopy.
NSTL
Elsevier