Measurement of feature clustering in images

Abstract

The spatial distribution of features in an image is often interesting, but not simple to characterize. Mapping of the image into a different space (e.g., Fourier or Hough) offers direct information on various regularities in feature spacing or alignment, but does not deal directly with the individual features. Two other approaches are available; each has advantages and drawbacks, which are discussed here.Schwarz & Exner determine the spatial coordinates of the centroids of features, and sort them to locate the nearest neighbor for each feature present, constructing a distribution plot of the frequency of nearest neighbor distances. Figures 1 and 2 show an example. The three fields in Figure 1 contain, respectively, features which are well-spaced from each other, randomly arranged on the plane, and clustered together. For the random distribution of points, the histogram of nearest neighbor distances is a Poisson distribution, and the mean value is 0.5/NA1/2, where NA is the number of features divided by the area of the image.