There is much information contained in the microstructure of a steel material. By quantifying the relative share of each phase and their morphology, it is possible to construct structure-property relationships and understand the overall material behaviour.

However, measuring microstructural characteristics is far from straightforward. The most intuitive approach requires identification of each pixel in terms of different constituents, such as phases, grain boundaries and defects, and quantification of their main morphological parameters. Unfortunately, manual labelling is extremely laborious and dedicated advanced imaging techniques are too expensive for routine use.

We therefore present two alternative routes based on deep learning to retrieve quantified microstructural information at a reduced cost. A first method extends on intuitive pixel-by-pixel annotation, speeding up the labelling by means of machine learning-driven segmentation methods. It retains metallurgical intuition, but it takes much effort to generate the required training data. A second path exploits advancements in neural networks and encoders to transform the images into low-dimensional vectorial representations, also referred to as latent features. These latent features lack direct physical meaning, but they are more robust in terms of training data and require less supervision during training.

We compare the efficiency of both types of microstructural features as input to structure-property relationships. The comparison is illustrated for steel grade classification and hardness prediction of generic binary Fe-C steels with a complex martensitic structure. The performance of latent features is found to outperform that of physical ones, even when using networks not specifically trained for handling microstructural images. It is a perfect example of how embracing computer vision can benefit the materials industry and does not always have to come with daunting learning curves.