Moment-Based Techniques for Image Retrieval

In this paper we analyze some shape-based image retrieval methods which use different types of geometric and algebraic moments and Fourier descriptors. Moments have been widely used in pattern recognition applications to describe the geometrical characteristics of different objects. They provide fundamental geometric properties (e.g. area, centroid, moment of inertia, etc.). We consider various description techniques: Hu, Flusser and Taubin invariants, Legendre and Zernike moments, Generic Fourier Descriptors (GFD). The set of absolute orthogonal (i.e. rotation) moment invariants defined by Hu can be used for scale, position, and rotation invariant pattern identification. Flusser´ s complete set of invariants appears as a particular case, with invariance only to rotation. The Taubin´s affine moment invariants introduce the concept of covariant matrix. Legendre moments are based on orthogonal Legendre polynomials and are not invariant under image rotation. Zernike moments consist of a set of complex polynomials that form a complete orthogonal set over the interior of the unit circle. GFDs are derived by applying a modified polar Fourier transform on shape image. We have applied the retrieval methods on a collection of images chosen from MPEG7 database. The image retrieval performance of each method is described by the precision-recall graph. In the paper we propose a novel approach that combines the described techniques after a coarse partitioning of the image dataset by their morphological features. The proposed approach provides much better performance than each method described above.