Images of the Russian Empire

Algorithm

I implemented the basic multiscale algorithm with the suggested vector dot product alignment metric, with a few modifications. Rather than operating over the entire image and using circshift to shift the layers, I operate only over the center 70% of the image and use my own shift function which just crops the two images being shifted on opposite ends. This prevents shifted portions of the image from interfering with the match. I use the scaled vector dot product alignment metric because it is both insensitive to overall changes in image intensity and image size. This is important because the pixels removed during shifting must not interfere with the alignment scores. I also added a variable width search window at the smallest level because the actual image size at that point can vary within a power of 2.

Results

Standard Images

The full results of my algorithm are available in my standard images directory. I have included a few samples below. I particularly like the last image shown, though it is apparent that pure shifting was not enough to completely align it's layers.

Additional Images

I ran a couple suitable images from a past photography assignment through my image alignment algorithm. The full results are available in my additional images directory. These images were taken using 35mm black and white film with three color filters. Aside from the modern camera and film format, the process is essentially exactly what Prokudin-Gorskii used.