Image Segmentation and Analysis Methods and Their Evaluation on Synthesized Porous Media Data
AdvisorRodriguez, Jeffrey J.
MetadataShow full item record
PublisherThe University of Arizona.
RightsCopyright © is held by the author. Digital access to this material is made possible by the University Libraries, University of Arizona. Further transmission, reproduction, presentation (such as public display or performance) of protected items is prohibited except with permission of the author.
EmbargoRelease after 07/10/2020
AbstractNondestructive imaging techniques such as X-ray computed tomography (CT) provide powerful means for quantitative analysis of porous material properties. Three-dimensional reconstructions from segmented X-ray CT images yield detailed information about density distribution, pore structure, pore connectivity, and phase interfaces that can be applied as boundary conditions for fluid dynamics simulations. However, accurate segmentation of grayscale X-ray CT data to discern solid medium constituents and fluid phases remains a significant challenge. To advance image segmentation, one objective of the dissertation research was the development of a new semi-automated multiphase segmentation algorithm combining K-means (KM) clustering with a Markov random field (MRF) framework. X-ray CT data were segmented with the new KM-MRF algorithm and with KM clustering only. A comparison of segmentation results shows that in the presence of noise inherent to X-ray CT data acquisition, KM-MRF yields fewer misclassification errors than sole KM clustering. Because the exact phase (i.e., solid, liquid, and air) boundaries of an imaged porous medium are not known a priori, there is no reliable reference data for meaningful validation of porous media segmentation algorithms. To overcome this problem, a second objective of the dissertation research was to synthesize a three-phase porous medium proxy with exactly known phase boundaries by using a discrete element method in conjunction with lattice Boltzmann fluid dynamics simulation. This approach generates an artificial porous medium with known phase boundaries, comprising spherical particles along with liquid and air. Poisson noise was added,and the contrast and resolution of the synthesized medium were varied to simulate image degradation experienced during X-ray CT data acquisition. The degraded data were then used to compare the performance of the KM-MRF, KM clustering, multi-Otsu and multi-SVM segmentation algorithms. The Dice and Jaccard similarity coefficients, and the misclassification, volume fraction, and surface area errors were used as performance criteria. The final objective of the dissertation research was the development of an efficient algorithm for quantification of phase interfacial area, a governing property for many porous media processes related to contaminant transport and remediation. An improved surface area estimator for three-dimensional objects based on the local Gaussian curvature was developed and compared with state-of-the art weighted-voxel techniques for five basic geometries. The relative error for the newly developed method was significantly smaller than the errors obtained with the competing weighted-voxel methods for objects with a combination of planar and a small proportion of curved surfaces, and comparable for objects with at least some proportion of curved surfaces.
Degree ProgramGraduate College
Electrical & Computer Engineering