Phase Classification by Mean Shift Clustering of Multispectral Materials Images

Abstract A mean-shift clustering (MSC) algorithm is introduced as a valuable alternative to perform materials phase classification from multispectral images. As opposed to other multivariate statistical techniques, such as factor analysis or principal component analysis (PCA), clustering techniques directly assign a class label to each pixel, so that their outputs are phase segmented images, i.e., there is no need for an additional segmentation algorithm. On the other hand, as compared to other clustering procedures and classification methods, such as segmentation by thresholding of multiple spectral components, MSC has the advantages of not requiring previous knowledge of the number of data clusters and not assuming any shape for these clusters, i.e., neither the number nor the composition of the phases must be previously known. This makes MSC a particularly useful tool for exploratory research, assisting phase identification of unknown samples. Visualization and interpretation of the results are also simplified, since the information content of the output image does not depend on the particular choice of the content of the color channels. We applied MSC to the analysis of two sets of X-ray maps acquired in scanning electron microscopes equipped with energy-dispersive detection systems. Our results indicate that MSC is capable of detecting additional phases, not clearly identified through PCA or multiple thresholding, with a very low empirical reject rate.