Report 5: handwritten character recognition

In this project, our goal is to define a set of features that allows us to carve up the feature space into regions that correspond to the character classes.

With the features implemented so far, it seems that just using "ink" and "log aspect", we can almost perfectly separate

The rest of the characters seem to be in a big blob in all coordinate-plane projections for the features we've implemented so far. Your task: try to devise some additional features so that as many as possible of the character classes can be separated by curves drawn in 2D coordinate-plane projections.

feature_space_carving.png

7:00pm, 10/26/18: I have manually erased extraneous dots in quite a few images which would otherwise have messed up some of the features, especially width, height, aspect. So please delete your image collection and redownload the updated ones: http://blue.math.buffalo.edu/448/handwriting/pngs.zip.

There are still some completely blank images. It would be nice to throw them out completely instead of leaving all their features at 0.