Automated Classification Of Scanning Electron Microscope Particle Images Using Morphological Analysis

We are developing a software tool that can automatically classify anthropogenic and natural aerosol particulates using morphological analysis. Our method was developed using SEM (background and secondary electron) images of single particles. Particle silhouettes are detected and converted into polygons using Intel's OpenCV image processing library. Our analysis then proceeds independently for the two kinds of images. Analysis of secondary images concerns itself solely with the silhouette and seeks to quantify its shape and roughness. Traversing the polygon with spline interpolation, we uniformly sample k(s), the signed curvature of the silhouette's path as a function of distance along the perimeter s. k(s) is invariant under rotation and translation. The power spectrum of k(s) qualitatively shows both shape and roughness: more power at low frequencies indicates variation in shape; more power at higher frequencies indicates a rougher silhouette. We present a series of filters (low-, band-, and high-pass) which we convolve with k(s) to yield a set of parameters that characterize the shape and roughness numerically. Analysis of backscatter images focuses on the (visual) texture, which is the result of both composition and geometry. Using the silhouette as a boundary, we compute the variogram, a statistical measure of inter-pixel covariance as a function of distance. Variograms take on characteristic curves, which we fit with a heuristic, asymptotic function that uses a small set of parameters. The combination of silhouette and variogram fit parameters forms the basis of a multidimensional classification space whose dimensionality we may reduce by principal component analysis and whose region boundaries allow us to classify new particles. This analysis is performed without a priori knowledge of other physical, chemical, or climatic properties. The method will be adapted to multi-particulate images.