Reliable Multi-Modal Automatic Remote Sensing Image Registration

In this thesis, both intensity and feature based automatic image registration approaches are addressed. In the first contribution of this thesis, the performance of both mutual information (MI) and cross-cumulative residual entropy (CCRE) as similarity measures are investigated in intensity based remote sensing image registration with Newton's gradient descent optimization algorithm. Although MI is the most popular choice for a similarity measure, the recently introduced CCRE measure was shown to be better than MI for multi-modal medical image registration. A Parzen-window approximation is applied to MI and CCRE to calculate their approximate gradients. Experimental results show that, for the case of registering several pairs of multi-modal images, including SAR, with a variety of affine deformations, CCRE provides superior performance to MI in terms of accuracy and success rate.In the second contribution of this thesis, a novel extension to the Parzen-window approximated maximization of similarity measures (MI & CCRE) is proposed which involves applying partial volume interpolation in the calculation of the gradients of the similarity measure. A new efficient implementation is also proposed which not only improves the computation time of MI and CCRE with partial volume interpolation but also improves the computation time of MI and CCRE. It is shown that introducing partial volume interpolation improves the performance of both CCRE and MI in terms of registration robustness and accuracy on the same set of data that was used to compare the performance of CCRE and MI.In the third contribution of this thesis, several issues of the scale invariant feature transform (SIFT), which is the most popular feature based image registration method, with multi-modal remote sensing images are addressed and their possible solutions are proposed. As the SIFT method is not designed for multi-modal images, it suffers very poor performance when registering multi-modal remote sensing images. The proposed modifications, collectively called MM-SIFT, have improved the reliability and robustness of the SIFT method for a variety of geometrical deformations (e.g., rotation, shear, scale, etc.) with multi-modal remote sensing images by a big margin. This improvement is verified on four sets of multi-modal images each containing several spectral bands