Machine learning is a growing topic with possibilities that seems endless with growing areas of applications. The field of metallography today is highly dependent on the operators’ knowledge and technical equipment to perform segmentation and analysis of the microstructure. Having expert dependents is both costly and very time-consuming. Some automatic segmentation is possible using SEM but not for all materials and only having to depend on one machine will create a bottleneck. In this thesis, a traditional supervised machine learning model has been built with a Random Forest (RF) classifier. The model performs automatic segmentation of complex microstructure features from images taken using light optical- and scanning electron microscopes. Two types of material, High-Strength-Low-Alloy (HSLA) steel with in-grain carbides and grain boundary carbides, and nitrocarburized steel with different amounts of porosity were analyzed in this work. Using a bank of feature extractors together with labeled ground truth data one model for each material was trained and used for the segmentation of new data. The model trained for the HSLA steel was able to effectively segment and analyze the carbides with a small amount of training. The model could separate the two types of carbides which is not possible with traditional thresholding. However, the model trained on nitrocarburized steel showcased difficulties in detecting the porosity. The result was however improved with a different approach to the labeling. The result implies that further development can be made to improve the model.
Keywords:
Machine learning; Metallography; Automatic segmentation; Complex microstructures; Random Forest classifier