Multi-Object Shape Retrieval Using Curvature Trees

摘要: With the increasing number of images generated every day, textual annotation becomes impractical and inefficient. Thus, content-based image retrieval (CBIR) has received considerable interest in recent years. For comparing images, CBIR uses generic features which are traditionally either intensity-based (color texture) or geometrybased (shape topology); latter is generally less developed than former. A common limitation existing geometry-based systems not considering simultaneously both shape topology objects (or components) may reveal important properties scene being analyzed. This work presents a approach for multi-object images. We commence with developing an effective matching method closed boundaries. Then, structured representation, called curvature tree (CT), introduced to extend handle containing multiple possible holes. also propose algorithm, based on Gestalt principles, detect extract high-level boundaries envelopes), evolve as result spatial arrangement group objects. At first, using triangle-area representation (TAR) presented non-rigid shapes The TAR 2D matrix that utilizes areas triangles formed by boundary points measure convexity/concavity each point at different scales triangle side lengths). capturing local global characteristics shape, invariant translation, rotation, scaling shear, robust against noise moderate amounts occlusion. matching, two algorithms introduced. first algorithm matches concavity maxima extracted from obtained thresholding TAR. In second dynamic space warping (DSW) employed search efficiently optimal (least cost) correspondence between shapes. dissimilarity derived correspondence. Experimental results MPEG-7 CE-1 database 1400 show superiority our over