2D-3D image registration of scanning electron microscope images and micro-CT volumes

Abstract

This thesis investigates the employment of feature based methods for 2D-3D image registration. Image registration of 2D images and 3D volumes has many industrial and medical applications. Industrial applications include grain structure analysis, data fusion for fault detection, medical image analysis. In this thesis we are interested in Micro-CT volume and SEM (Scanning Electron Microscope) image registration. State of the art 2D-3D registration techniques are computationally expensive. For example, registration of a Mega-pixel 2D image to a Giga-voxel 3D volume computation time is in many hundreds of CPU hours utilizing state of the art CPU technology. We address this problem by introducing a point feature based registration framework for 2D-3D registration. Specifically, we present novel techniques for repeatable key-point detection for 2D images and 3D volumes, a novel way of descriptor formation for matching 2D key-points and 3D key-points and model estimation for registration up to an affinity under low feature matching accuracy. We further develop techniques for model estimation. We present an affine registration estimation based on an algorithm, which re{u00AD} quires three true positive matches of feature points and extend this to estimate the model based on only a single true positive match, which we call the one{u00AD} point algorithm. We present an algorithm based on Branch-and-Bound for rigid model estimation with proof that convergence is guaranteed. We experimentally validate the performance of this algorithm and show its theoretical superiority compared to RANSAC. In addition, handling image scale is also addressed by resolving the ambiguity of 2D image scale and 3D image scale. We show that 2D scale and 3D scale do not represent similar image and volume neighborhoods. We compare our technique with state of the art global registration techniques, such as correlation based registration and mutual information based registration and indicate the superior performance of our method. Furthermore, we introduce a novel feature descriptor based on image curvature founded on mathematically sound principles for improving feature matching accuracy. We indicate that the 2D-3D registration accuracy improves under this novel feature descriptor. Other practical applications, such as homography estimation and pose estimation, are also investigated. Furthermore, we extend the Branch-and-Bound based algo rithm with guaranteed convergence for vanishing point estimation and essential matrix estimation, for which empirical results are provided. Show more