Three-dimensional shape characterization for particle aggregates using multiple projective representations

Abstract

Shape descriptors are used extensively in computer vision/automated recognition applications such as fingerprint matching, robotics, character recognition, etc. The conventional two-dimensional shape descriptors used in these applications do not readily lend themselves to compact representations in three dimensions. The situation is even more challenging when one attempts to numerically describe the three-dimensional shapes of a mixture of objects such as in an aggregate mix. The goal of this study is the design, development and validation of automated image processing algorithms that can estimate three-dimensional shape-descriptors for particle aggregates. The thesis demonstrates that a single set of numbers representing a composite three-dimensional shape can be used to characterize all the varying three-dimensional shapes of similar particles in an aggregate mix. The composite shape is obtained by subdividing the problem into a judicious combination of simple techniques - two-dimensional shape description using Fourier and/or invariant moment descriptors, feature extraction using principal component analysis, statistical modeling.and projective reconstruction. The algorithms developed in this thesis are applied for describing the three-dimensional shapes of particle aggregates in sand mixes. Geomaterial response such as shear strength is significantly affected by particle shape - and a numerical description of shape allows for calculation of functional characteristics using other previously established models. Results demonstrating the consistency, separability and uniqueness of the three-dimensional shape descriptor algorithms are presented.