In this project, we took glass plate images produced by Prokudin-Gorskii and digitized by the Library of Congress and processed them into colored images. Prokudin-Gorskii took three pictures of all his subjects, one with a red filter, one with a green filter, and one with blue filter. My goal was to pick a color channel and align the other two color channel images with that one. Then, I would combine all three aligned color channels to produce a colored image. For smaller jpg images, I used an exhausted search over a [-15, 15] window in both the x and y direction, searching for the optimal offset combination (x, y) that produced the lowest cost based on a loss function. I implemented Normalized Cross Correlation and Sum of Squared Differences as the two loss functions but use the latter. The offset chosen produced the lowest SSD. For larger tiff images, I used a pyramid search since an exhaustive search would be too computationally expensive since we'd need to search over a larger space. In this approach, I start with the smallest scale image, performed an exhaustive search over the [-15, 15] window. Then, I scale the image and the offsets up by 5/3s. We do this recursively until we reach the original resolution. This allows us to cover a wider search space.
At first, the pictures did not align very well using the exhaustive search. After aligning it with the green channel instead of the blue channel, the results improvee signifcantly. For the larger tiff images, the pyramid search was initially far too slow and the images were blurry. However, suprisingly, cropping a huge portion of the image (500 pixels on each side) both improved results and sped up the search. This is likely because the most important information is encoded in the internal pixels closer to the center
blue-green offset: (-2, -5), red-green offset: (1, 7), Cathedral
blue-green offset: (-4, -25), red-green offset: (-8, 33), Church
blue-green offset: (-24, -48), red-green offset: (17, 57), Emir
blue-green offset: (-17, -59), red-green offset: (-2, 64), Harvesters
blue-green offset: (-17, -41), red-green offset: (5, 49), Icon
blue-green offset: (-8, -51), red-green offset: (3, 60), Lady
blue-green offset: (-10, -83), red-green offset: (4, 101), Melons
blue-green offset: (-2, 3), red-green offset: (1, 6), Monastery
blue-green offset: (-27, -50), red-green offset: (10, 58), Onion Church
blue-green offset: (-29, -79), red-green offset: (9, 103), Self Portrait
blue-green offset: (-15, -50), red-green offset: (-2, 59), Three Generations
blue-green offset: (-3, -3), red-green offset: (1, 4), Tobolsk
blue-green offset: (-6, -41), red-green offset: (26, 43), Train
blue-green offset: (1, -53), red-green offset: (-11, 52), Workshop
