Date of Award

December 2016

Degree Type

Dissertation

Degree Name

Doctor of Philosophy

Department

Engineering

First Advisor

Zeyun Yu

Second Advisor

Roshan M. D'Souza

Abstract

Image matching techniques are proven to be necessary in various fields of science and engineering, with many new methods and applications introduced over the years. In this PhD thesis, several computational image matching methods are introduced and investigated for improving the analysis of various biomedical image data. These improvements include the use of matching techniques for enhancing visualization of cross-sectional imaging modalities such as Computed Tomography (CT) and Magnetic Resonance Imaging (MRI), denoising of retinal Optical Coherence Tomography (OCT), and high quality 3D reconstruction of surfaces from Scanning Electron Microscope (SEM) images. This work greatly improves the process of data interpretation of image data with far reaching consequences for basic sciences research. The thesis starts with a general notion of the problem of image matching followed by an overview of the topics covered in the thesis. This is followed by introduction and investigation of several applications of image matching/registration in biomdecial image processing: a) registration-based slice interpolation, b) fast mesh-based deformable image registration and c) use of simultaneous rigid registration and Robust Principal Component Analysis (RPCA) for speckle noise reduction of retinal OCT images. Moving towards a different notion of image matching/correspondence, the problem of view synthesis and 3D reconstruction, with a focus on 3D reconstruction of microscopic samples from 2D images captured by SEM, is considered next. Starting from sparse feature-based matching techniques, an extensive analysis is provided for using several well-known feature detector/descriptor techniques, namely ORB, BRIEF, SURF and SIFT, for the problem of multi-view 3D reconstruction. This chapter contains qualitative and quantitative comparisons in order to reveal the shortcomings of the sparse feature-based techniques. This is followed by introduction of a novel framework using sparse-dense matching/correspondence for high quality 3D reconstruction of SEM images. As will be shown, the proposed framework results in better reconstructions when compared with state-of-the-art sparse-feature based techniques. Even though the proposed framework produces satisfactory results, there is room for improvements. These improvements become more necessary when dealing with higher complexity microscopic samples imaged by SEM as well as in cases with large displacements between corresponding points in micrographs. Therefore, based on the proposed framework, a new approach is proposed for high quality 3D reconstruction of microscopic samples. While in case of having simpler microscopic samples the performance of the two proposed techniques are comparable, the new technique results in more truthful reconstruction of highly complex samples. The thesis is concluded with an overview of the thesis and also pointers regarding future directions of the research using both multi-view and photometric techniques for 3D reconstruction of SEM images.

Share

COinS