Phase preserving Fourier descriptor for shape-based image retrieval

Abstract

Shape is one of the most important discriminative elements for the content based image retrieval and the most challenging for quantification and description. Fourier descriptors are a very efficient shape description method used in shape-based image retrieval tasks. In order to achieve invariance under rotation and starting point change, most Fourier descriptor implementations disregard the phase of Fourier coefficients, consequently losing valuable information about the shape. This paper proposes a novel method of extracting Fourier descriptors that preserve the phase of Fourier coefficients. We introduce specific points, called pseudomirror points, and use them as a shape orientation reference. They facilitate the extraction of phase-preserving Fourier descriptors which are invariant under translation, scaling, rotation and starting point change. The proposed descriptor was tested on four popular benchmarking datasets: MPEG7 CE-1 Set B, Swedish leaf, ETH-80 and Kimia99 datasets. Performance and computational complexity measures indicate that the proposed method outperforms other state-of-the-art phase-based Fourier descriptors. In addition, it outperforms other state-of-the-art magnitude-based Fourier descriptors, and many non-Fourier based shape description methods in terms of performance – complexity ratio.