Statistical Extraction of Meaningful Descriptors for Object Shapes
John C. Russ, PhD, Professor, North Carolina State University
F. Brent Neal, PhD, Research Scientist, Milliken & Company
Shape is an elusive concept. Humans recognize many different things in their world based on the presence or absence of key features, which they also must learn to recognize. Teaching computers to classify objects based on shape, or to measure shape for correlation with history or performance, is a classically difficult task. It is important in forensics, natural sciences, engineering, and even "simple" tasks such as optical character recognition for reading printed texts. It may either be concerned with identifying a class of objects, a specific individual or a distinctive mark. A wide range of mathematical measurement tools exist that can characterize certain aspects of shape, including dimensionless ratios, Fourier or wavelet coefficients, invariant moments and more. But determining which of these is most appropriate or useful in a given instance generally requires collecting a set of training data and exploring it statistically. A variety of specific applications, including natural and man-made objects, are used to illustrate the process, using discriminant analysis and neural nets.