15-463 Computational
Photography
Project 1: Colorizing
The Prokudin-Gorskii
Photo Collection
<>Lindley French
lms@andrew.cmu.edu
I originally attempted to use Sum Squared Error to align the various
color channels. However, this proved ineffective on a number of images.
Therefore, I switched over to Normal Correlation. I was able to further
improve my results by only correlating the inner 3/5ths of the image in
each dimension. I made use of an image pyramid to increase processing
speed on large images, so that at each level of the pyramid, only 25
variations needed to be considered. It may be possible to reduce this
number further, but I felt it was better to err on the side of caution.
I attempted to implement a means by which the meaningless color-bands
on the borders of the images were automatically cropped, by testing the
normal correlation of individual rows and columns. My results in this
area were not ideal, although they did reduce the banded area
significantly. The main problem I faced in this regard was finding
threshold values which cropped most of the bands, without needlessly
cropping the actual image.
There is one image which my algorithm did not handle well without
special tweaking. More on that below.
The first images were medium-quality jpeg files, and were processed
quickly.
The channels did not completely align in this image, but I suspect that
is due at least in part to wind.
The below image, unlike all the other images on this page, did not
produce usable results with my program. I was able
to obtain the below so-so result by subtracting from each image itself
circshifted by [-1 -1], but this modification was not ideal,
and it produced bad results on other images. Therefore the modification
in question is only used for this files with the same
name as this one. It is not an elegant situation, but it was the best I
could manage in the time allotted.
The below images were originally high-quality tif files. I have shrunk
them and converted to jpeg for presentation here.
The high-quality versions are linked from the smaller versions.
