Project 1: Colorizing photos

Brandon Huang, Fall 2017

Project overview

In this project, we restored the color of images from the Prokudin-Gorskii collection, a collection of glass plates exposed in 3 colors that records color photographs. In order to do so, we needed to align the 3 filtered channels correctly, as the alignment is slightly off to begin with. For my approach, I used normalized cross-correlation (NCC) as a metric to compare color channels at different offsets, under the (slightly flawed) assumption that the exposures in the 3 channels would be of similar distribution. I implemented image pyramid for the higher resolution TIF images to reduce the number of NCC computations needed to arrive at a precise alignment.

Challenges

Initially I used the SSD metric to compare images, but NCC proved to be a more accurate metric. Then, I had trouble tuning the parameters of the image pyramid. The optimal depth and crop amount ended up being 4 and 200 respectively; it was also helpful to search more extensively at the deepest level, when the NCC operation is the cheapest.

Colorized images

Blue offset: (-5, -2); Red offset: (7, 1)

Blue offset: (-49, -24); Red offset: (59, 17)

Blue offset: (-60, -16); Red offset: (66, -3)

Blue offset: (-40, -18); Red offset: (49, 5)

Blue offset: (-54, -8); Red offset: (63, 3)>

Blue offset: (3, -2); Red offset: (6, 1)

Blue offset: (-3, -1); Red offset: (4, -1)

Blue offset: (-79, -29); Red offset: (99, 7)

Blue offset: (-7, 0); Red offset: (8, -1)

Blue offset: (-53, -14); Red offset: (59, -4)

Blue offset: (-42, -7); Red offset: (44, 26)

Blue offset: (-56, -21); Red offset: (60, 8)

Blue offset: (-65, -13); Red offset: (73, 9)

Other colorized images from the collection

Blue offset: (-26, 19); Red offset: (90, -18)

Blue offset: (-9, -4); Red offset: (19, -9)

Blue offset: (-24, -20); Red offset: (36, 8)

Blue offset: (-12, -6); Red offset: (77, 9)