Machine-Learning-Based Image Similarity Analysis for Use in Materials Characterization

Abstract

Materials with similar microstructural features exhibit similar properties, a fact which often provides useful insights for a detailed understanding of the materials. An analysis of material similarity in terms of microstructural images is proposed for predicting some properties of interest. This similarity analysis is inspired by the application of medical image retrieval to guide diagnostic decisions. Some relevant analyzing techniques including machine-learning algorithms of zero-normalized cross-correlation, mutual information, maximum likelihood estimation, principal component analysis, and self-organizing map are applied in this work. These techniques are systematically employed to identify the variances of the query images based on the metrics of image natural properties (such as brightness which is measured in pixel) or metallurgical features contained in the microstructural images. It is shown that the employed methods exhibit consistent similarity evaluation results. The proposed similarity analysis of microstructural images is expected to provide a new avenue for understanding the materials paradigm.