Dissertation Defense of Ben Glocker

Title: Random Fields for Image Registration

Abstract:

Image registration is one of the key components in computer vision and medical image analysis. Motion compensation, multi-modal fusion, atlas matching, image stitching, or optical flow estimation are only some of the applications where efficient registration methods are needed. The task of registration is to recover a spatial transformation which aligns corresponding structures visible in the images. This is commonly formulated as an optimization problem based on an objective function which evaluates the quality of a transformation with respect to the image data and some prior information. So far, mainly classical continuous methods have been considered for the critical part of optimization.

In this thesis, discrete labeling of random fields is introduced as a novel promising and powerful alternative. A general framework is derived which allows to represent both linear and non-linear image registration as labeling problems where random variables play the role of transformation parameters. Based on this framework, several explicit models are defined for the linear and non-linear case. While discrete optimization often provides strong solutions in purely discrete settings, the task of registration actually involves the estimation of continuous transformation parameters. In order to bridge this gap, a novel optimization procedure is proposed based on iterative discrete labeling with successive label space refinement strategies. The procedure is computationally efficient, avoids local minima through large neighborhood search, and yields high-accurate registration.

Besides efficiency, the great advantage of this discrete formulation is that it provides an intuitive control on the search and solution space, prior knowledge can be easily integrated, and it is modular in terms of the objective function since neither numerical nor analytical differentiation is necessary. The implementations are based on the most recent advances in discrete optimization. Performance of the methods is evaluated in numerous medical and non-medical applications such as multi-modal registration, segmentation via atlas matching, deformable image stitching, and optical flow. Experimental results show consistently the great potential of random fields for image registration. This thesis aims at creating a novel and valuable perspective on the modeling part also for other imaging and vision tasks, and hopefully influences the way people think about optimization and the applicability of discrete random fields beyond classical problems.

Committee Head: Prof. Struss, TUM
1st Referee: Prof. Navab, TUM
2nd Referee: Prof. Paragios, ECP, France

Date: Monday, June 9, 2011

Time: 11:00

Location: MI 01.09.014

Technische Universität München
Fakultät für Informatik
Boltzmannstr. 3
85748 Garching b. München
Germany