# Random Image Generation using Iterated Function Systems

I became, like many, interested in fractals and computer image generation after reading Chaos, James Gleick's 1987 popular science classic. One part of that book that really caught my attention was connected with the picture shown at right. This image of a fern leaf was generated using Michael Barnsley's method of iterated function systems. What was so amazing about this was that despite the manifest complexity and detail of the image, a natural shape, it took only a few parameters to specify it -- just 4 sets of six real numbers, 24 all together, to 2 digits of precision. 48 decimal digits: 160 bits.

Here, I thought, was something quite amazing. Perhaps equally remarkable was the way that the numbers were used to generate the image, which seemed almost like magic. Each set of 6 numbers specifies a function known as an affine transformation from the plane to the plane. Such a function takes points in the plane (or shapes, which are sets of points) and maps them to rotated, skewed, translated versions of themselves. (Technically, it consists of a linear transformation plus an added translation.) To generate the image, one picks any point close to the origin and randomly applies one of the six transformations to it. The result is plotted, and then another randomly chosen transformation is applied to this point, that result is plotted, and so on. (This is the origin of the name iterated function system.) Somehow, from this sequence of random choices, the image-pattern emerges, independent of which point is started from, or what sequence of choices is made.

Before I learned anything further about how (and why) this actually works, I wondered what would happen if instead of using Barnsley's fern leaf parameters, you substituted in ones that had been generated randomly. My (faulty) reasoning was that the IFS technique mapped a small space of parameters onto a much larger space of images, and since the result in one case was known to be aesthetically interesting, there was reason to suspect that the subspace mapped onto consisted entirely of interesting images. With breathless anticipation I watched as my old 8086 IBM clone ground out the points of my first experiments in this area -- which turned out to be utterly disapppointing random collections of dots and lines.

Undeterred (and spurred on by my younger brother's derisive reaction to my attempts), I went back and learned a little more about how iterated function systems worked and used this insight to constrain the random parameter generation a little bit. The results (see examples below), were quite rewarding. It is possible to generate pleasing abstract shapes via random parameter generation, the characteristics of which depend on the constraints employed. Also, by slightly modifying the parameters for a given shape, one can obtain interesting variations of it (the two images in second row are distortions of Barnsley's fern).

What are these images? Plants? Clouds? Snowflakes? They seem to define an entire realm of quasi-natural shapes. Almost every new image is a pleasant surprise.

The images were generated by a program I've written called randim, which affords interactive control over the parameter selection process. (Some more recent images, with color: one two and three. Screenshots of the program itself in action here.) You can download this program if you want (HTTP). It is written in C using X Windows and should work on most workstations with minor adjustments at most. All I ask is that you don't sell the program or one based on it for money, and you send me the code for any cool modifications you do (email: arobert at interstitiality ddot net). Thanks!

## References

A technical explanation of how iterated function systems work (along with a brief description of my randomization constraint methods) can be found here.

A clearer and more rigorous explanation can be found in Barnsley's excellent book, Fractals Everywhere, cited below.

Practically, iterated function systems have been applied to image compression and font generation, and theoretically they have relations to dynamical systems and neural networks. Personally, I think there is much more that remains to be said about them. I continue to be amazed at the sheer variety of patterns that can be generated via random parameter selection. Somehow the mathematical structure of the IFS technique opens the connection between randomness and regularity, chaos and order, the absence of symmetry and its presence - that appears to be a central aspect of natural beauty.

1. Barnsley, M.F. (1988); Fractals Everywhere.
2. Barnsley, M.F., Hurd, L. P. (1993); Fractal Image Compression.
3. Ip, H.S., Wong, H.T.F., Mong, F.Y. (1994); "Fractal Coding of Chinese Scalable Calligraphic Fonts" in Computers and Graphics 18:343-51.
4. Stark, Jaroslav (1991); "Iterated Function Systems as Neural Networks" in Neural Networks 4:679-90.