Visual Pathways for Shape Abstraction

摘要

The Medial Axis Transform (MAT) (or skeleton transform) is one of the most studied shape representation techniques with established advantages for general 2D shape recognition. Embedding local boundary information in the skeleton, in particular, has been shown to improve 2D shape recognition capability to state of the art levels. In this paper we present a visual pathway for extracting an analogous to the MAT skeleton abstraction of shape that also contains local boundary curvature information. We refer to this structure with the term curvature-skeleton. The proposed architecture is inspired by the biological findings regarding the cortical neurons of the visual cortex and their special purpose Receptive Fields (RFs). Points of high curvature are initially identified and subsequently combined by means of a visual pathway that achieves an analogous to the MAT abstraction of shape but also embeds in the skeleton local curvature information of the shape's boundary. We present experimental results illustrating that such an abstraction can improve the recognition capability of multi layered neural network classifiers.