Skeletons from dot patterns: A neural network approach

Abstract Boundary detection is a well studied problem in the context of shape extraction from dot patterns and digital images. For images, particularly binary images, another frequently encountered issue is finding the skeleton of the object. Unfortunately, in the case of dot patterns, the skeletonization problem has not received much attention due to the lack of a proper definition of a dot pattern skeleton. We present a method, using artificial neural networks, to extract the skeletal shape of a dot pattern and demonstrate that the skeleton thus obtained is close to the perceptual skeleton. The neural network model proposed here is a modified version of Kohonen's self-organizing model. It is dynamic in the sense that processors can be inserted (or deleted) during the learning process. Unlike in Kohonen's map, the number of processors here need not be known a priori.