Not much here yet. I waste my time experimenting with segmenting images.

Why? Image segmentation is really Big AI: no amount of playing with edge detectors, low pass filters, and Gabor transforms will result in decent image segmentation (i.e. understanding the scene's elements). It is fundamentally a high level recognition problem. Hard is fun.

I've been playing with Kohonen networks. Here's the output from a 2 layer network. It was presented with a few jpgs that I grabbed from the net. The first layer broke the pictures into regions (using hue and eight intensity filters.) The second layer tried to extract common shapes from the resulting regions. This pic is the archetypes chosen by the second layer. The bowtie shapes are an artifact caused by the limited angular resolution of the shape layer.