Classification of graphite particles in metallographic images of cast irons – Quantitative Image Analysis versus Deep Learning

Characterizing graphite shape in cast irons is of importance for control of cast components. This is still done based on standard charts and has been more and more supported by quantitative image analysis during the last decades. Despite the significant progresses made by image analysis techniques and software packages, the results of quantitative image analysis are still moderately satisfactory, and the question raised of the capabilities of machine learning to supplement this approach.

In this work, a first set of challenging images -in which both lamellar, compacted and spheroidal graphite coexist- was submitted to standard image analysis and to a convolutional neural network, a deep learning model. The comparison of the two methods shows that machine learning produced very encouraging results when compared to standard image analysis, in particular for what concerns the presence of lamellar graphite.

Then, an extended series of micrographs in which degenerate graphite shapes were included was processed with an improved version of the neural network. After appropriate training, the results were again satisfactory.