The goal of this project is to be able to align split images from the Prokudin-Gorskii collection. The spite images are split into Blue, Green, and Red channels and must be aligned to produce a full RGB image. Throughout the course of the project, I explore two main approaches to complete this task: exhaustive search and image pyramid search. Out of all the images, there is also an exception with some images where we adjust our method to align the image.
For our initial exhaustive search, the Blue channel is the reference channel and the Red and Green channels iterate over the offset range of [-15,15]. Essentially, the R and G channels are experimentally aligned over the B channel to find the alignment with the lowest error. To calculate the error, when comparing displacements, we cropped the RGB channels by the max displacement value (15) and compared B with R and B with G using their norms. The examples below also had 15% of the edges cropped in order to remove the black border for better accuracy.
Cathedral
G: [5, 2], R: [12, 3]
Monastery
G: [-3, 2], R: [3, 2]
Tobolsk
G: [3, 2], R: [6, 3]
Since TIFFS are much larger images, exploring the same displacements as the exhaustive search would take too much time. The pyramid search helps by downscaling the image, finding a rough displacement, then aligning and finding a slightly less rough displacement on a lesser downscale until we go up all the way to the full-scale image. Using this method, we don't need to check as many pixels at the largest size. For my process, I downscale by 0.5 until the image reaches a size of 100 pixels. The displacement achieved at the downscaled image is multiplied by 2 when aligning the image at the larger level and continues until the full-scale. My process was also done recursively and 15% of edges were cropped to help with accuracy.
Lady
G: [50, 9], R: [111, 12]
Onion Church
G: [51, 26], R: [108, 36]
Church
G: [25, 4], R: [58, 5]
Melons
G: [81, 9], R: [177, 12]
Harvesters
G: [59, 16], R: [123, 13]
Icon
G: [40, 17], R: [89, 11]
Self Portrait
G: [77, 28], R: [174, 36]
Three Generations
G: [52, 12], R: [110, 10]
Train
G: [42, 5], R: [86, 31]
Workshop
G: [52, -1], R: [103, -12]
Emir was a bit special due to all the blue that is in his clothing. Using G and R referenced over B was not effective and gave poor results. So, we switched the process to B and R referenced over G to achieve the below results.
Emir
B: [-49, -24], R: [57, 17]
Tree
G: [53, 24], R: [117, 33]
House
G: [45, 6], R: [97, 3]
