dc.description.abstract |
This thesis presents a detailed study and improvement to local descriptor processes for
registering images for the purpose of three-dimensional reconstruction, using four M¯aori
artefacts as case studies. The motivation for the research came from the issues which still
exist in image registration when dealing with large magnitudes of image transformations.
Four major pieces of work were carried out in the course of this research. First, an
evaluation was carried out to study the performance of local descriptor processes and based
on the results, the local descriptor process was divided into three stages, of which two
were closely analysed. Second, the local descriptor formation stage was studied, and two
methods, colour and hybrid local descriptor methods, were developed using colour images
instead of greyscale images to improve the uniqueness of local descriptors. Third, the local
descriptor matching stage was studied, and a new method based on support vector machines
was developed. Fourth, an assisted image registration programme was developed and is a
semi-automatic approach for registering images.
Extensive amount of experiments were carried out to validate these work. It was found
that the colour and hybrid local descriptor methods had gains in matching accuracy of
up to 10% over existing methods, and the support vector machine maching method had
increased matching performance of up to 20%. When the two methods were combined, it
was found that performance gains of up to 25% could be achieved. For the assisted image
registration programme, up to 50% improvement was achieved, and the advantage was more
significant as the magnitude of image transformation increased, highlighting the need for
such programme.
These results show that the proposed work in this research are significant contributions to
literature. In addition, these results show that the proposed methods can be used successfully
for registering images for three-dimensional reconstruction, where the image transformation
between images are often large. As there is currently a need to reconstruct M¯aori artefacts,
this research has provided a new approach for registering images of these artefacts, which
could then be used to construct three-dimensional models of the artefacts. |
en |