The goal of this project is to take a glass plate image and reconstruct it as a colorized image.
Approach
I decided to implement the Sum of Squared Differences method, and tested it on both the .jpg and the .tif images.
The .tif images took significantly longer than the .jpg to process, and results varied. Some, like tobolsk, came out fine, whereas others, like self_portrait, did not align properly.
I then added the pyramid speedup as well user input for pyramid level and scale, and ran the colorizer on all the images at various pyramid level counts and scales.
I started to see better results overall, but some, particularly harvesters, emir, and train still had very noticable alignment issues.
Next, I manually added cropping, cropping off 10 pixels per side for .jpg and 100 pixels per side for .tif. This only saw some improvement in alignment, and the errors were still very noticable.
Increasing the cropping on .tif images to 150 on each side drastically reduced the blurring on emir, harvesters, self_portrait, and train. The black borders significantly affected the aligning of some of the images, therefore removing them improved the alignment quality.
For some images like train, 6 pyramid levels at scale 0.5 and 150 cropped off each side yielded better results than the same level.scale with no cropping or 10 pyramid levels at scale 0.75 and 150 cropped off each side
Examples
Offset given in (axis0, axis1)
10 level images slightly shrunken to meet size limitations.
6 levels, 0.5 scale, No Cropping
10 levels, 0.75 scale, Crop 10(jpg), 150(tif)
Self-picked
All of the following images colorized with 6 levels, 0.5 scale, and 150 pixels cropped off each side.