A study on robust feature representations for grain density estimates in austenitic steel

Modern material sciences and manufacturing techniques allow us to create alloys that help shape our way of living; from jet turbines that withstand extreme stresses to railroad tracks that retain their intended shape. It is therefore an important aspect of quality control to estimate the microstructural properties of steel during and after the manufacturing process, as these microstructures determine the mechanical properties of steel. This estimation has for a long time been a labor intensive and non-trivial task which requires years of expertise.
We show that modern deep neural networks can be used to estimate the grain density of austenitic steel, while also applying a visualization technique adapted to our task to allow for the visual inspection of why certain decisions were made. We compare classification and regression models for this specific task, and show that the learned feature representations are vastly different, which might have implications for other tasks that can be solved via discretization into a classification problem or treating it as an estimation of a continuous variable.