Connected component labeling using self-organizing feature maps

A three-phase algorithm is presented for labeling the connected components of a binary image. The importance of the algorithm is that a connectionist parallel implementation will be proportional only to the number of points in the image, and only two parameters need to be tuned, one directly controlling the number of iterations and one to determine the desired number of connected components. The approach taken is to consider the object points in the image as cities in the traveling salesman problem and then use the optimization power of self-organizing feature maps to find a near-optimum path. It is found that this path can be partitioned into a number of subpaths, each representing a connected component. Theoretical work is needed to support these experimental findings.<>