Icon Shuffle

I have a large collection of icons and thought it would be interesting to see if the software would be able to detect similarity between them. The next picture shows you a subset of my icons and you can see some of them are quite similar. Click it to enlarge.



I've written some code to compare all icons to each other and the next picture shows you how the icons are paired. The right icon of each pair is the best found match for the left icon.



Overall the matching is quite okay. In most cases it finds a variant of an icon as its best match. Sometimes the best match is weird, because there is a variant but it is not selected. For some reason the red heart is chosen as best match in a number of cases.

Sometimes the best match is really another object but we can understand the similarity. For example this one where the flow of the object probably caused the match:


Or this one where the colours and form match quite well:


And the roundness of these two objects:


Okay, I'll be moving on to other data sources from now on. I found a site where I can easily retrieve stock market data and feed it into the Crazy Diamonds engine. Again I'm going to search for similarity but now for stock value history. Maybe we'll come up with some interesting reoccuring patterns there. I'll let you know soon!

First Contact

I've been tweaking my Crazy algorithms for the last week with ups and downs but finally a configuration came out that I'm quite happy with for the moment. I've been using 2D images to test my software and in this post I want to show you some results. It is not spectacular at all (yet) but it shows that base system is able to recognise patterns in different images. So let's take a look at some examples:

First I drawed a very schematic toucan (3 ellipsoids) and compared that with one of my Vista backgrounds (toucan in a forest). The software scans the background image and searches 3 spots that match my schematic drawing as good as possible. Here we have the input images and the results:


We see the 3 spots the system found. The result images are grey because we only visualise the luminance feature. The top image is the best match, the bottom image the worst of the 3. Although our query image was very schematic it is able to find two acceptable results. The third result is just plain rubbish but I didn't analyse why this spot came up in the matching. Two out of three ain't bad said Meatloaf.

In my next test I put a bit more effort in drawing the toucan by hand to see if the matching result would improve considerably. And guess what!

Yep now it was able to really pinpoint the toucan in the picture. The best result is on the left, the worst on the right. The best two results really match the drawn toucan very well. I felt like a happy bunny...

Not all tests were successfull. For example, I tried to find a flower bud in a field of flowers and it couldn't find it unless I tweaked the parameters to make it find it. But overall I was quite pleased with the results. Finally I decided it was time for an Eye Test:

I asked the Crazy system to find me some letters and I'll show you the results for searching 'A' here:

This result really thrilled me because Crazy didn't only find the A instance I was looking for (top result) but also its little brother (bottom result)! Here we can see that wavelet decomposition extracts pattern data on multiple resolutions and therefore find the smaller 'A'. Cool!

I can show you more pictures, but you probably get the idea from the examples above. I'll come back to you soon with new Crazy results, bye bye for now.

Let's get technical

Okay, let's have a look at our technical design.


The picture above shows the main modules of our pattern extraction / matching system:

The Source Linking module links binary data sources such as images, music/sounds, texts and stock trading data to the system.

The Pattern Analysis module extracts characteristic patterns from the data source using wavelet decomposition techniques.

The Pattern Storage module stores all pattern information in a datastore. Each data source consists of a set of features, such as luminance and colour intensity for images.

The most interesting part is the Pattern Matching module that is able to calculate the similarity of patterns. The matching function is based on the Earth Mover's Distance algorithm and is able to compare n-dimensional patterns and therefore capable to match patterns for a range of different types of data sources (e.g. stock exchange data, images, sound, 3D flight patterns etc.)

Next time I''l show some first results of the system in action.

Introduction

In this blog I will be presenting new results of my ongoing "Crazy Diamonds" project. This 'research' project assumes that everything in our universe can be expressed as a pattern interfering with other patterns. Interfering patterns generate new patterns. Every pattern needs a medium (carrier) to express itself and we introduce 'Consciousness 'as this underlying medium. Consciousness is the living essence of all that is, not observable because it is the observing itself. I told you this was going to be easy.

I have a more detailed project introduction on my website, but for now it is only available in dutch:

http://www.indranet.nl/crazydiamonds1.htm

If you would like to have a summary in English, please let me know. Other suggestions, reactions are always welcome!

The project does not have a clear goal in mind, but it will focus on how we can present things as patterns and how these patterns relate to each other. Maybe, just maybe, I will find out how weather changes in the Middle-East are affecting the stock market in New York, in which case I will be swimming in money very soon :-) Naturally you can start buying "Crazy Diamonds" stocks right now, just send me a note.