Shape-based image retrieval for JPEG-2000 compressed image databases

As JPEG-2000 is primarily based on wavelet transforms, we propose a moment-based algorithm to extract shape features out of wavelet coefficients to facilitate image retrieval from compressed image databases. Motivated by the observation that isolated significant points in the wavelet coefficients domain cannot accurately represent the shapes of regions, a range of ideas are investigated in this paper to construct shape descriptors in the wavelet domain using several morphological operators. Based on the fact that certain moments are invariant to rotation, a more compact and effective moment representation method is also proposed to integrate shape information in various directional wavelet sub-bands. Extensive experiments have been carried out using JPEG-2000 compressed image databases, and the results show that the proposed algorithm outperforms a representative existing shape-based image retrieval method in the wavelet domain.