The main task of forming technology is the customized modification of the shape and properties of metallic materials. The microstructural characterization of the materials is very important and also recognized, whereby their classification is not a trivial task. Classification appears to be extremely difficult, especially when mixtures of different phases with different substructures are involved. So far, there is no sophisticated computer system that allows an automatic classification of microstructures.
Therefore, the main objective of our research is an attempt to apply techniques from the field of image analysis based on Deep Convolutional Neural Networks (DCNN) in segmenting different types of microstructures.
The core of our research was to establish the interaction between database creation of microstructure images, DCNN architecture development and materials science expertise, and to evaluate the system.
In particular, by successfully applying unsupervised learning methods, it was possible to train the system without predefined target values or other influences.
With 1.2 million images of steel structures, our image database is one of the most extensive data collections worldwide.
The obtained results confirm that the system based on a DCNN algorithm coupled with machine learning methods can be useful in segmenting and identifying undefined microstructure types of steels, although the architecture of the network needs to be continuously developed for more accurate prediction.