Statistical modeling in wavelet domain for Bayesian texture classification and retrieval

Feature extraction and similarity measure in feature space are two basic steps in a texture based image retrieval system. In this paper, we propose to extract statistical modelbased features in wavelet domain where we use dual tree complex wavelet transform (DT-CWT). To this end, we employ generalized Gaussian density (GGD) to describe the statistical characteristics of DT-CWT coefficients. On the other hand, we utilize Bayesian classifier for measuring similarity and so texture classification. In addition, for improving the classification rate and computational complexity we project the features onto a low dimensional space using three methods: linear discriminate analysis (LDA), locality preserving projections (LPP) and kernel LDA (KLDA). Our experiments are conducted on two different texture databases, i.e. VisTex and Brodatz. We achieve the classification rates up to 97.5% and 96.54% for these two databases respectively which validate the robustness of the proposed method.