My algorithm used a simple sum of squared differences to align the images. I experimented briefly with other metrics, such as NCC, but found that using a different metric did not significantly affect image alignment quality. Instead, I worked to optimize the search process with a simple metric. My recursive search, using the pyramid algorithm, cut each image down to 250 pixels in size before searching for a good alignment. The number of possible alignments tried at each level of recursion scaled inversely with the size of the image, so that a reasonable amount of time was spent at each level of the recursion.
For those images which did not align well, I think that a more sophisticated metric would be necessary, but I wasn't able to find any that worked significantly better.
Higher-resolution images can be found here.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|