Using fractals to learn image descriptions by means of artificial neural networks

Compressing an image for later reconstruction requires encoding, even with fractals, a large amount of information. On the other hand, if the goal is not to reproduce with high fidelity an image, but to distinguish between instances of different classes of them, it seems reasonable to think that the amount of information to be extracted from a single image may be drastically reduced, provided that one has a method for capturing the characterizing “essence” of the image. The authors´ claim is that, for at least some classes of images, this method is provided by the fractal reconstruction of the image. The authors present a set of experiments, aimed at investigating the previous claim. As images the authors have selected m=9 classes of trees. Each prototype has been generated with a single fractal rule, which, starting from a base shape, appends to the growing figure smaller and smaller copies of itself