CS 194 Project 1

Sergey Mikhaylovich Prokudin-Gorsky (August 30, 1863 – September 27, 1944) was a Russian chemist and photographer. He is best known for pioneering colour photography and his effort to document early 20th-century Russia. He travelled across Russia in a railroad-car darkroom using three-image color photography to record the things he saw. He envisioned classrooms would have special projectors that would display his RGB negatives so he could bring color photographs to Russian children.

In this project, we are given the three exposures (RGB glass plates) for various photographs and asked to optimally align them to recover a single, colored photograph.

Main Project

Initial Approach

As described in the project spec, my initial, naive approach was to extract the individual RGB channels, calculate the SSDs produced by shifting the R and G photos through a [-15, 15] window, and align the RG channels to the B channel based on the minimum SSD found. However, in photos where one color was more predominant, the SSDs would be skewed and result in a misaligned photo.

Improvements

The following are improvements I made to my algorithm to improve the photo alignment (excluding bells and whistles):

Started the aligning process by cropping the borders of each photo by 10% to get rid of any skew caused by the dark colors.
On larger photos, the naive method of searching a [-15, 15] shift window resulted in 5+ minutes of runtime. I implemented the image pyramid method where I rescaled the photo by a factor of 1/2 in each recursive call until I reached the base case (photo dimensions < 250 pixels). From there, I would run the [-15, 15] window-shifting approach, then pass the offsets up through the recursive calls, checking a [-1, 1] window with each pass to fine tune the coarse-grain offset found at the base case.

Bells and Whistles

Better Features

I used the skimage.filters.roberts method to implement edge detection and remove bias in alignment that results from dominant colors.

On most of the images provided in class, the difference between what the image looks like with or without edge detection is not super significant. Since everyone is using Emir as an example, I'll show the result of what edge detection looks like on one of the extra photos I picked from the collection.