An effective technique for multi-modal image registration

Abstract

Image registration brings two images into alignment despite any initial misalignment. Multi-modal image registration is a special type of registration problem where complexity is added, due to images being captured by different types of imaging devices. For decades, researchers have continued to work on multi-modal image registration. Yet limited attention has been paid to developing local feature descriptors for multi-modal registration, despite their useful performance in general registration problems. We would argue that multi-modal registration can therefore benefit from qualities that are inherent in local techniques. In this thesis, we focus on developing a new multi-modal image registration technique based on local descriptors. We study a range of local feature description techniques, acquire an in-depth understanding of relevant sub-procedures and propose a novel modality invariant feature description technique that can be effectively applied to multi-modal image registration. The proposed technique has four components which are carefully designed to cater for variations in image properties caused by multi-modality as well as to enable finding correspondences with high accuracy under a wide range of transformations. We identify certain properties of multi-modal images, namely gradient and region reversal, which must be addressed appropriately in order to register multi-modal images. The first component of our proposed technique overcomes problems with existing local multi-modal techniques in addressing these properties. The second component comprises strategies to encode gradient profiles that are appropriate for multi-modal registration applications. The third component employs a robust and powerful data fitting technique, namely Hough Transform, in order to achieve high accuracy in multi-modal registration. Our proposed solution is simple, but effective, and enables seamless integration of Hough Transform by overcoming existing technical limitations. The final component incorporates special methodologies to cater for local rotations in input images. Besides evaluating our technique in terms of the accuracy of identified correspondences, we also show that the proposed technique results in less registration error when compared with existing techniques.