Rotation-invariant Bivariate features for texture image retrieval

Abstract

Considering the inter-scale dependency between the coefficients, a novel progressive rotation-invariant texture retrieval means based on inter-scale dependency is proposed in this paper. Firstly, Logpolar transform and Non-Subsampled Contourlet Transform (NSCT) are combined to get rotation-invariant multi-scale and multi-direction coefficients, then Generalized Gaussian Distribution (GGD) model is used to extract the profile information from low band which could be employed further as coarse retrieval features. Afterwards, the inter-scale dependency is modeled by Non Gaussian Bivariate Model and is used as fine retrieval foundations. Experiments on Brodatz standard texture database show that, our method provides better efficiency and accuracy with lower feature dimension compared to wavelet transform and intra-scale model GGD and is proved to be an efficient rotation-invariant texture retrieval means.