The purpose of this project was to utilzie image processing tequniques on digitized images from the Mikhailovich Produkin-Gorskii photo collection to output colored images. By seperating three image channels (R, G, and B) which were obtained by taking three photos in quick successing using red, green, and blue plates, three images were obtained. I then used a Sum of Sqaured Differences (SSD) alignment ttequnique over each image to identify the best possible alignment of the three color channels. I aligned the blue and red channels using the green channel as the reference because the image resutls (particularly for emir) were most accurate with that approach.
The resulting images were all consistenly of similarly high quality after some experimentation with window sizing, border adjucement and the parameters for the recursive pyramid algorithm.
First, I removed 8% of the borders of each image. 8% was selected based on experimentation.
Then, I used a window size of 4 and increased it in increments of 3 in my recusrive pyramid algorithm as the resulution was halved at each iteration. This allowed me to look over ranges of up to 40 pixels for the lowest quality images, as scale back down to 4 at the final iteration.
For images smaller than 250X250 pixels, I used a default window size of 15x15 and set the initial x and y shifts to (0,0)
For the smaller images, the naive algorithm was run achieving the following offsets:
For the largest images, the pyramid algorithm was run recursively achieving the following offsets:
In addition to the images provided, I run my algorithm on a couple additional images found on the website of the Prokudin-Gorskii collection. Below are the results for these images as well
The most difficult images proved to be ones with many colors and intricate detail. I had a a lot of trouble with emir and village in particular. One key reveltaiton that helped me make a lot of forward progress was to switch from using the blue channel as the reference to the green channel. I think that makes sense and increases the overall quality since green is the middle channel so aligning the top channel and bottom channel to it makes sense.