Unfortunately, simply setting each of the three images Gorskii took to red, green, blue channels produces bad results because the images are not quite correctly aligned (see left). Correctly aligning images, however, yields much better results. I aligned images using the following methods:

• First, I applied a canny edge filter to each channel to create an image that would be easier to align. I chose a canny edge filter because this filter produced much better results than the more widely known sobel filter.
• Second, I found a good alignment of the R and B channels with B by finding the translation that minimized the sum squared error.
• For images with less than 100,000 pixels I searched for translations from -15 to 15 in x,y. For images larger than this I used a Gaussian pyramid to recursively downsample the image by 1/2 the size until reaching a small enough image, then recursively adjusting the alignment over a window of -4 and 4 in x, y.
• Finally, I cropped 10% of the image around the edge of the image to eliminate edge effects (however, some are still visible).

Using this approach most of my images aligned quite well. However two of the larger images (see below, larger images), 0087 and 01403 still has some misalignment artifacts. I resolved this for 01403 by aligning to the G channel instead of the B channel. For 0087, I was still unable to find a good alignment, I think this is because the subject actually moved when the photo was taken.

Overview of Results

The tables below illustrate results for small files as well as larger files. Each image also shows the offsets I used to align the images. In addition to showing results required for class, I have included two results of my choice in both the small and large set.