A Novel Shape Based Hierarchical Retrieval System for 2D Images

Shape is an important property of visual objects which distinguishes it from others and is done with ease by living beings. Shape is invariant to scale (magnification or shrinking), rotation, translation, and flip (mirror-reflection). A good representation will lead to a better classification or retrieval system. A known representation called “normalized centroid distance signature” has all above properties except invariance to rotation and flip. By quantizing the signature into “signature histogram”, invariance to rotation and flip are also achieved. But, this has a severe drawback, namely, number of false positives will increase, i.e., many entirely different shapes can have same or similar signature histograms. The paper proposes to augment the signature histogram with local information that is present at the boundary of the object. This is done in such a way without loosing any properties of the signature histogram. This leads to a hierarchical representation scheme called the “kth order augmented histogram” where k is a positive integer. A novel kth order hierarchical retrieval system is also proposed which uses the hierarchical representation scheme. The proposed scheme is compared with a recent similar scheme.