Fractal Based Unstructured Object Recognition

Many conventional methods of object recognition depend solely on the
geometric structure of the underlying objects. This strategy is
eminently suitable for man-made, or robot-made objects.
Natural objects are, for the most part, unstructured
in the sense of Euclidean geometry. But there is
an inbuilt degree of self-similarity in the image of natural objects
which can be efficiently utilized through a representation exploiting
self-transformations,
known as Iterated Function
System (IFS).

Interestingly, virtually all images of natural {\em or} man-made objects, show
region-wise self-similarity although they may not be globally
self-similar. Such objects can be represented by Partitioned Iterated
Function System (PIFS) very compactly.

In this paper we propose an object recognition scheme based on such
PIFS representation and matching carried out in the PIFS code
domain, which we argue is more efficient than correlation in the image
domain. We base our proposition on theoretical arguments.
The recognition method is attractive from the
view-point of execution time as the PIFS code library of reference
objects are built off-line and recognition of query object involves
only comparison of its PIFS code with those in the library.

Keywords: Object recognition, geometric structure, unstructured object,
self similarity, fractal, Partitioned Iterated Function System.