| dc.description.abstract | Hyperspectral (HSI) and Multispectral (MSI) images consist of multiple images taken at different wavelengths (bands), often at different times. Analysis of these images require high accuracy of correlation between the different image bands. As the image bands are taken at different times, different positions, or both, this entails that the images must be processed before the images can be analysed. The process of alignment between two images is referred to as image registration.
A test-set consisting of 40 image bands was created, consisting of wavelengths in the range of (415-557nm). These images are 1601x1401 pixels, and have introduced simulating errors, to be corrected by the registration process. This enables validation of the registration results. Insight Toolkit (ITK), a C++ library, was explored and implemented for this test-set. ITK was chosen, as this library contains most known registration methods, and boast high customizability. Six registration methods were implemented and tested against this image set.
One of these registration methods, Rigid transform, was found to achieve sub-pixel precision across multiple input parameters, with 0.22 standard deviation from the true pixel coordinates, and 0.018 standard deviation from the true angular value. This method has an optimal run time of six hours per image band using one central processing unit (CPU) core. The remaining five registration methods showed promise, but were not reliable against the test-set. These registration methods ranged between 33 and 76 standard deviation from the true pixel coordinates.
Rigid transform was then run on the unknown hyperspectral and multispectral image sets. Quantifiable accuracy results are not available for these image sets. | |
