CS 129 Project 1 Writeup

Charles Yeh (ccyeh)
September 16, 2012

In order to quickly test multiple alignments, my project creates an image pyramid for the three different images being aligned, and matches them at smaller dimensions because matching them at higher ones. The image pyramid is created by blurring, sampling every other pixel, and repeating for a given number of levels. This allows my project to a few possibilities at each dimension to narrow options down, rather than checking all possibilities right away.

The image pyramid had little effect on small images, but a rather large effect on the tif’s which were very large.

I compared sum of squared differences, and also normalized cross correlation, but they were pretty much the same results-wise.
See last set of images for the final output of my project.

Extra credit:

• I used sobel edge detection for better image matching. Using edge detection allowed the images of the different color channels to be more similar, and the alignments were generally better. Some images were aligned the same.

Original images:
With edge detection:

• I also did automatic cropping. I do this by taking the minimum area of the aligned images, and trimming down the sides 2 lines at a time, stopping when the standard deviation (of the two lines about to be removed) is greater than a threshold currently set at .18. The problem with having a set threshold was that lower contrast images would always have lower standard deviations. I didn’t have the time to have my project calculate a threshold from the image’s contrast but that’s the next step I would’ve done. Instead, I set a limit as to how much could be cropped.

Uncropped images:
Cropped images: