Analysis and visualization of volumetric data sets.

摘要

The ability to generate volumetric data sets of materials, particularly at sub-millimeter resolutions, has increased in recent years. The tools to facilitate analysis of non-medical volumes are still primitive and require substantial user input. This thesis addresses three tasks of volumetric analysis, producing efficient algorithms for segmentation, interest detection, and view path generation.;The first major contribution is a clustering based segmentation algorithm that was designed to operate on noisy volumetric data. The initial data is a reconstruction from tomography or MRI and is inherently noisy. The segmentation algorithm operates by detecting statistically similar, although possibly non-contiguous, regions and then relabeling the volume monotonically using a Bayesian classifier based on the detected clusters in the volume.;The second major contribution is the development of a trainable interest detector. A Radial Mass Transform (RMT) is defined to characterize the local structure at each location in the volume. This transform is demonstrated to provide a rich characterization of local structure. The RMT is used as the base input to a Support Vector Machine (SVM) classifier to generate an application specific interest detector. The user selects a set of interesting and uninteresting regions in the volume, the SVM is trained using the labeled data, and the entire volume is classified.;The third major contribution is a view path, or tour, generating algorithm. Once volumes have been processed, a visual representation of the volume is helpful for users to better understand the structure of the data. An algorithm is developed to automatically generate view paths that maximize the amount of interest shown in a series of cross sections. The cross sections can be viewed as an animation, or virtual tour, providing the user with additional information about the structure of object(s) in the volume.;The algorithms are demonstrated on hundreds of synthetic volumes and dozens of real volumes, including many microvolumes scanned using the Argonne National Laboratory APS. Two common threads running through this work are that all techniques were designed to work on large volumes and to operate in parallel when possible. A typical data set is on the order of 400--600 megabytes and may require several hours, if not days, of processing time on a desktop computer. The techniques we have developed are amenable to parallel processing, reducing processing time from days to hours or minutes.