Probabilistic complex phase representation objective function for multimodal image registration

Abstract

An interesting problem in computer vision is that of image registration, which plays an important role in many vision-based recognition and motion analysis applications. Of particular interest among data registration problems are multimodal image registration problems, where the image data sets are acquired using different imaging modalities. There are several important issues that make real-world multimodal registration a difficult problem to solve. First, images are often characterized by illumination and contrast non-uniformities. Such image non-uniformities result in local minima along the convergence plane that make it difficult for local optimization schemes to converge to the correct solution. Second, real-world images are often contaminated with signal noise, making the extraction of meaningful features for comparison purposes difficult to accomplish. Third, feature space differences make performing direct comparisons between the different data sets with a reasonable level of accuracy a challenging problem. Finally, solving the multimodal registration problem can be computationally expensive for large images. This thesis presents a probabilistic complex phase representation (PCPR) objective function for registering images acquired using different imaging modalities. A probabilistic multi-scale approach is introduced to create image representations based on local phase relationships extracted using complex wavelets. An objective function is introduced for assessing the alignment between the images based on a Geman-McClure error distribution model between the probabilistic complex phase representations of the images. Experimental results show that the proposed PCPR objective function can provide improved registration accuracies when compared to existing objective functions.