Groupwise image alignment automatically provides non-rigid registration across a set of images. Groupwise non-rigid image alignment is a difficult non-linear optimisation problem, involving many parameters and often large datasets. It has found applications in facial image analysis and medical image analysis by automatically generating statistical models of shape and appearance. The main approaches used previously include iterative and manifold learning-based approaches. In iterative approaches, the registration of each image is iteratively updated to minimise an error measure across the set. Various metrics and optimisation strategies have been proposed to achieve this, but requiring complex optimisation, often with many unintuitive parameters that require careful tuning for each dataset. In this thesis, we restructure the problem to use a simpler, iterative optimisation algorithm, with very few free parameters. We demonstrate how to incorporate a stiffness constraint and how to tune the few remaining parameters. The warps of training images are refined using an iterative Levenberg-Marquardt minimisation to the mean, based on updating the locations of a small number of points and incorporating a stiffness constraint. The use of LevenbergMarquardt optimisation is made possible by using a small number of points at each iteration, as it would be too expensive to use with many parameters. It also has the advantage of stability and is self tuning.
|Date of Award||2020|
|Supervisor||Bernie Tiddeman (Supervisor)|