CS 180 Portfolio

Project 1: Colorizing Prokudin-Gorskii Photos

Overview

Sergei Mikhailovich Prokudin-Gorskii (1863-1944) recorded three exposures of every scene onto a glass plate using a red, a green, and a blue filter. Our goal is to take the digitzed Prokudin-Gorksii glass plate images and produce a color image.

Approach

After separating the three different images, the first thing I did to measure the similarity between two different pictures was to just take the Euclidean distance between the two. For the green and red images, I would shift the x and y coordinates by [-20, 20], and recalculate the Euclidean distances. This worked okay, not super great, so I decided to try and crop the sides by 5%. This would help alleviate the issues with the potential black edges messing up the Euclidean distance, or the next metric I tried, NCC. Instead of actually improving it, cropping made it worse. So, I decided to try a new metric -- using Sobel edge detection. This automatically improved the alignments quite a bit, so this is the method I went with. Then, the next step was to process the .tif files. I had tried these when running the Euclidean distances, but the file size was too large and took too long. So instead, I decided to recursively minimize the quality of the image, search for the alignment there, and then scale it back up. I settled for a pixel size of 400, as I found that it worked the best. Then, as I scale the image back up by 2x each time, I only need to search a radius of [-2, 2] in both the x and y direction after scaling back up, since we found the best alignment from the scaled down version of the picture. This is the final method I ended up with, as well as cropping the sides by 15%, as I found that worked best as well.