Dr Lakaemper,

 

Please find attached my HW2 solution (please rename the file into .zip and then unzip it).

 

The programs that you should run are part1.m and part2.m .

 

Other two *.m files contain functions.

 

Query function takes an additional parameter, which is a cell array with images, and it returns indices to the two most similar images (not the actual images).

 

Histogram similarity is based on the norm of histogram difference. However, before that I apply histogram normalization (so that the sum of a histogram is always 1) and smoothing (with a sliding window of size 11). Smoothing eliminates (or decreases) the possibility of situation where two images that look very similar can have very different histograms, for example one image is obtained from the other by adding a constant small value to all RGB values.

 

I added some comments to the *.m files, but I did that on a computer with no Image Processing Toolbox, so I cannot test the programs again. I hope I did not introduce any errors into the source code.

 

I include 8 images (4 pairs) and a separate database of images (in ‘source’ folder) that I collected from Wikipedia “Featured Picture” archive.

 

When selecting 4 pairs of images, I “cheated” by using a program that created a pairwise similarity matrix for all images in ‘source’ folder. I selected 4 pairs of images that have the best similarity out of all pairs. However, it turns out that some of these pairs are also very close to each other. When running part2.m, for some images the two most similar images are not from the same pair. Better classes and representatives could be obtained by using clustering (into 4 clusters), and then selecting representatives that are closest to the cluster centroids.

 

Uros Midic

uros@ist.temple.edu

Phone +215.204.5908

FAX +215.204.5082

Temple University, 303A Wachman Hall, 1805 N. Broad St., Philadelphia, PA 19122