Skeletonization of sparse shapes using dynamic competitive neural networks

Abstract

The detection of regions and objects in digital images is a topic of utmost importance for solving several problems related to the area of pattern recognition. In this direction, skeletonization algorithms are a widely used tool since they allow us to reduce the quantity of available data, easing the detection of characteristics for their recognition and classification. In addition, this transformation of the original data in its essential characteristics eases the elimination of local noise which is present in the data input. This paper proposes a new skeletonization strategy applicable to sparse images from a competitive, dynamic neural network trained with the AVGSOM method. The strategy developed in this paper determines the arc making up the skeleton combining AVGSOM non-supervised learning with a minimum spanning tree. The proposed method has been applied in images with different spanning shape and degree. In particular, the results obtained have been compared to existing solutions, showing successful results. Finally, some conclusions, together with some future lines of work, are presented.