Shape Skeleton Extraction And Shape Decomposition Based On Visual Saliency Features

With the development of the Internet technology, the amount of image datahas become larger and larger. They contain abundant image information that we canutilize. How to obtain useful information has thereof triggered many activeresearch fields. As for shapes, shape features play an important role in humanvisual system, and can distinguish one object from the others. Shape analysis andrepresentation have been widely applied in object recognition, computer graphics,multimedia search, medical image analysis.Shape representation is the first step and most critical step towardsshape-related applications. However, shape representation faces big challenges,because the shapes of real objects vary differently and are usually influenced bynoise and segmentation errors. We aim to extract the most discriminative androbust features. The research in this paper concentrate on shape representation,where shape skeleton extraction and shape decomposition based on visual saliencyfeatures are two main aspects. The main contents and contributions are as follows:(1) We propose a novel skeleton growing algorithm combining the strength oftwo complementary shape saliency measures, i.e. discrete curve evolution (DCE)and bending potential ratio (BPR). We first analyze the weakness of the algorithmof skeleton pruning by DCE. Our improvements are twofold:1) The pruning is notdone in post processing (after the skeleton is computed), but is integrated into theskeleton growing process and the compact skeleton is directly obtained.2) Thepotential redundant skeleton branches are pre-constrained through introducing BPRconstraint to key skeleton points. The experimental results on noisy and deformedshapes demonstrate the valid of the proposed algorithm.(2) We propose a novel shape decomposition algorithm. In general, we utilizethe information provided by structural and boundary features, i.e. skeleton andintegral invariant. The novelty of proposed method has threefold:1) By combiningthe strengths of skeleton and integral invariant, we predict candidate split pointsfollowing rules of skeleton "junction points" and "the minima of contour curvature".2) We also propose a mechanism of avoiding conglutination of important parts byutilizing the mark information of shape contour.3) A bending potential ratio is introduced as a constraint to generate controllable decomposition results.Experiment results show that our method satisfies subjective visual perception onshape decomposition and is stable.