Assignment 1:
Images of the Russian Empire
The goal of this assignment is to take the digitized Prokudin-Gorskii glass plate images and, using image processing techniques, automatically produce a color image with as few visual artifacts as possible. In order to do this, I extracted the three color channel images, placed them on top of each other, and aligned them so that they form a single RGB color image. Below are a few samples of the original images that I received to colorize. All of the original images can be found here.
In general, the results are very clean. The artifacts on the edges are not cropped away so one can realize how significant an improper alignment can affect the outcome. The images were generated on a 2.8Ghz Core2Duo with 4GB of memory. The low-resolution images utilized a normalized 2D cross-correlation algorithm, while the high-resolution images utilized the NCC algorithm in the base case and a sum of squared differences algorithm at each succeeding step of the pyramid. The generation of the low-resolution images took very little time, a couple of seconds; the generation of the high-resolution images took approximately 30 seconds each.
Low-resolution single scale images:
There were 13 low resolution glass-plate images provided, which date back to the Russian empire...except one image of Hamerschlag Hall. :) These images were aligned using a normalized 2D cross-correlation algorithm, which performs pretty quickly. The resulting images after extraction and alignment are as follow:
High-resolution multi-scale images:
There were 7 high resolution images provided. These images were aligned using an image pyramid for optimization. At the base case, which I denoted as any image less than 1000x1000px, an NCC algorithm was used for accurate alignement estimation. At each step of the pyramid building up, an SSD algorithm was used since the SSD is able to specify exactly which offsets we want to score. As a result, the combination of both of these methods provided a fast and accurate colorization of the images. The following images are scaled. Full versions can be found here. The high resolution images are provided in both jpg and tif formats.
User-chosen images
The following images were chosen by me, with one that turned out very interesting. The image of the ocean sunset was the only image that I found which failed to align properly. This was likely due to the extreme differences in the images in the glass-plate. Below are some images that I liked:
Remarks
Overall, I was very satisfied with the quality of the output. Almost all of the images have minimal artificats due to alignment. One can still notice alignment issues if he looks carefully, though. An example is provided be below. We can see that there is a green outline on the picture frame (?), which is clearly due to a misalignment. While this artifact is minor, it shows that it takes a lot of work to restore images well.