Images of the Russian Empire

Aaron Johnson, Computational Photography Fall 2007 Proj1

Initiallly for this project I did a simple Sum of Squared Differences distance for image alignment. From this I added an image pyramid which reduced the area needed to search at each step, and thus the overall speed. To enhance the alignment, I made the assumption that while color values in different channels could be different, the edges would probably be the same across all colors. Using this thought I strengthened the edges before calculating the template matching. This helped to lign up images with strong edges, but not as much in images with weak images. The amount of edge strengthening can be varied accordingly. The last thing I did was a pretty simple algorithm to crop most of the ugly color bars that the shifting of images causes around the perimeter. The algorithm basically looks for the last column who's maximum value is less than a certain threshold for a given color value. This implies that that column is entirely low-energy, and probably not part of the image.

*It is interesting to note that 2 of the base set, and one of the additional images, did not align correctly with the edge enhancements on. After i reduced the weight or turned them off the alignment was fixed on all but the one that is shown misaligned. That image has a repeating pattern on the door that often gets "stuck" one row off from correct, throwing the whole image off. At the same time many images did not align correctly with them off. There may be a middle ground that gets all images, or a better edge enhancement algorithm.

Now, on to the images. First a zoom in on one of the high-res images when the algorithm was used with edge enhancement:


And now the same region after the image was alligned without the edge enhancement. Note the green tinge on the corner near the bolt, and the red/blue shift on the right/left sides of the tiles:


This is a more drastic change, first with edge detection:


And now without:


Here is an example output for an image with the borders cropped:


Note that there is still discoloration around the edge on this and other final images, but there is at least some data there so I did not crop it off. Here is that same output without the borders cut off:


I chose these images in addition because of the large amount of color in them:

*


And now onto the other images:
*



*