In this project we take scans made by Prokudin-Gorskii and bring them together into a single color image.
We divide the image into three parts, and then align the red and green color channels with the blue channel. The alignment is done by searching over cyclic shifts, done at progressively finer scales. In this way we are able to align images that have displacements on the order of hundreds of pixels.
The distance between the channels is computed as a frobenius inner product of the internal pixels. Using the internal pixels improved alignment greatly, as the signal from the inner product can get polluted by high distances between border regions.
Some images were sensitive to different hyperparameters in the alignment algorithm. These two images in particular are aligned correctly if we limit the coarse-graining threshold in the pyramid recursion. However, with settings that gave good results for the majority of the images, these two images are not properly aligned.
