Material phase classification by mears of Support Vector Machines.

Abstract

The pixel's classi_cation of images obtained from random heterogeneous materials is a relevant step to compute their physical properties, like E_ective Transport Coecients (ETC), during a characterization process as stochastic reconstruction. A bad classi_cation will impact on the computed properties; however, the literature on the topic discusses mainly the correlation functions or the properties formulae, giving little or no attention to the classi_cation; authors mention either the use of a threshold or, in few cases, the use of Otsu's method. This paper presents a classi_cation approach based on Support Vector Machines (SVM) and a comparison with the Otsu's-based approach, based on accuracy and precision. The data used for the SVM training are the key for a better classi_cation; these data are the grayscale value, the magnitude and direction of pixels gradient. For instance, in the case study, the accuracy of the pixel's classi_cation is 77.6% for the SVM method and 40.9% for Otsu's method. Finally, a discussion about the impact on the correlation functions is presented in order to show the bene_ts of the proposal.