Classifiers performance evaluation in quantitative metallography

As the need for increasing speed takes place at industrial processes in general, the use of digital techniques such as image processing and automatic classification have been playing an important role at materials characterization and quantitative metallography fields. The aim of this work was to develop and evaluate computational vision techniques on solving ordinary problems as the area fraction of phases determination of AISI 1020 steels. Three techniques were implemented and evaluated: k-Nearest Neighbors (KNN) which consists in classifying pixels based on their neighborhood information, Artificial Neural Networks and Support Vectors Machine, these last two centered on supervised machine learning processes. Indexes that denote classification quality were then evaluated. Concerning classification time and relative accuracy, the SVM results have shown superiority. Nevertheless, in all cases, the classification values have agreed with the area fraction values expected for this type of steel based on theoretical metallurgical analysis.

Image classifiers; Quantitative metallography, Image analysis, Microstructural characterization.