by: John Taylor Gatto

I just found out today that my mail server in VA was down.  Well, actually it wasn’t down exactly; it was trying to come back up.  My mail server’s spam filter didn’t start-up on reboot. So when there was a power failure a little while back, the machine just sat there rejecting any/all email.  It would have started up on its own but I had to disable windows server’s shutdown ‘reason’ tracking.  Anyway, I’ve done all that now, so the machine should be able to restart with email fully functional now…

Since I had nothing to do today (for once), I decided to fire up the old pressure cooker and cook up somthin’ home style - that’s guju home-style.

This pic shows the spices I use in my cooking.  If you don’t know what these spices are…go get some culture. ; )

I have been meaning to update my blog with the HAAR filter and other goodies, but I have been pretty bogged down at my new job.  As you can imagine, working at DreamWorks (not to mention relocating to a new place) leaves very little time for blogging.  But once things settle down, I’ll post the wavelet example and more…

In one of his talks, Richard Dawkins discusses an observation of guppies in insolated pools of water. In one pool there are predators with the guppies and in another there are no predators at all. Over the course of evolution the guppies in the pool without predators evolved to have different vibrant coloring; the guppies in the pool with predators, dull, muddied coloring.

The reason there is more variation in the first pool, Dawkins explains, is due to the fact that the selection pressures that would be introduced by a resident predator are not present; leaving the primary selection pressure, sexual selection, as the main evolutionary control. Female guppies choose male guppies based on their coloring (perhaps this is simply because of their greater visibility).  Vibrant coloring would have been selected out of the pool with predators as colorful guppies would be equally visible to mates as well as predators.

This just illustrates how over predation kills variation in prey.  You just end up with ugly guppies that taste good to insatiable predators.

Here you can see different results obtained from the Neural Edge Detector.  Underneath each image is the number of training cycles used to train the network.  It is pretty obvious from these results that increasing the number of training cycles results in a “more informed” determination of edge/non-edge pixels. 

I have created a simple example that you can run yourself.  In the .zip file below you will find a folder that containes the above images as well as the (win32) ProcessImages.exe application.  There is also a sub-folder which contains different network descriptions (layers/nodes/thresholds/weights).  To use this example, on the command line, you simply type the command name with the appropriate network file and image file:

ProcessImages <networkFile> <imageFile>

Example:

ProcessImages “NetworkFiles/SavedNetwork-10000-0.1.txt” originalsmall.jpg

File:

Training Tests

After implementing the Sobel and Laplace edge detection algorithms and building my neural network library, I thought it would be interesting to see if I could apply the pattern association aspect of neural networks to the edge detection problem.  It turns out to be an awesome idea.

I have written a small Image Processing Workbench that will allow me to easily explore different Image Processing algorithms.

I took this picture in South Africa a couple of years ago. It’s the best picture from the whole trip.