Nanodispersion, nonlinear image filtering, and materials classification

Polyethylene terephthalate - alumina nano-composites from two production processes gave rise to materials H and T, further divided into four and, respectively, three classes of belonging. Electron microscope images of the materials had been visually scored by an expert in terms of an index, β, aimed at assessing filler dispersion and distribution. These properties characterize the nano-composite. Herewith a classification algorithm which includes image spatial differentiation and non-linear filtering interlaced with multivariate statistics is applied to the same images of materials H and T. The classification algorithm depends on a few parameters, which are automatically determined by maximizing a figure of merit in the supervised training stage. The classifier output is a display on the plane of the first two principal components. By regressing the 1st principal component affinely against β a remarkable agreement is found between automated classification and visual scoring of material H. The regression result for material T is not significant, because the assigned classes reduce from 3 to 2, both by visual and automated scoring. The output from the non-linear image filter can be related to filler dispersion and distribution.