Analysis of Vector-Directed and Best-Parameter Search Algorithms for Transformation Function Construction In Automatic Image Registration
Date of Award
Doctor of Philosophy (PhD)
Benjamin Ray Seyfarth
Automatic image registration is in high demand due to the enormous quantity of satellite, medical, and aerial images. This dissertation aims to develop an automatic image registration approach to the registration of aerial images of the Mississippi Delta area. Since there are many significant line edges in the aerial images, a line edge detector, the local standard deviation (LSD) technique is designed, and it is shown to be a successful line edge detector. The similarity metric, normalized cross-correlation (inner dot product, specifically), is demonstrated to be a reliable similarity measurement for images processed by the LSD technique. The developed transformation function search strategy, the vector-directed search algorithm, takes the affine transformation as a 5-parameter-vector function and searches toward the gradient direction that results in higher similarity values. The conducted experiments have shown that combined with the LSD technique, the vector-directed search method can successfully identify an optimal transformation function, provided the search starting point is relatively close to the best transformation function. In addition, another search algorithm--the best-parameter search algorithm is introduced, and it has been demonstrated to be as successful as the vector-directed search algorithm. As the registration of the aerial images requires registering a sequence of overlapping images, the initial-overlap alignment technique is developed, and it successfully ensures the automatic registration of all the aerial images in a sequential order.
Zhou, Hong, "Analysis of Vector-Directed and Best-Parameter Search Algorithms for Transformation Function Construction In Automatic Image Registration" (2004). Dissertation Archive. 2697.