Skeletal Saliency Map Computation Based on Projection Symmetry Analysis

Abstract Skeleton is one of the simplest but powerful tools for shape representation and analysis, which encodes the original shape by one less dimension but can retain the topological information well. Inspired by the observation that the skeleton also serves as a symmetry axis, we propose to compute the skeleton by symmetry analysis of the given shape. In implementation, we define the skeletal saliency map by using the projection symmetry measurement. Firstly, the primitive boundary is extracted. Then, for each pixel p in the shape, we project its four neighboring pixels onto the boundary. The saliency of p is estimated by the maximum distance along the boundary between the two projections. Finally, a clean skeleton is reported by a simple truncation operation. The proposed method can produce a two-pixel-wide skeleton with correct topology and avoid any additional post-processing. Furthermore, it is also robust to boundary noise. Extensive experiment results validate these nice properties of the proposed method, which shows that it outperforms the state-of-the-art methods.