Polar-wavelet energy signatures for rotation-invariant texture classification

This paper proposes a scheme to extract rotation-invariant polar-wavelet energy signatures for texture classification. Firstly, we apply polar transform to a given texture and decompose the transformed image to generate a set of redundant shift-invariant sub-bands of wavelet packet coefficients with orthonormal wavelet bases. Then we compute the polar-wavelet energy signatures from each sub-band, and select the most dominant features to form a feature vector for texture classification. The experimental results, based on testing 7200 texture images, with different orientations, against 25 texture classes, show that the proposed scheme using the Euclidean classifier outperforms two other common rotation-invariant texture classification methods, its overall accuracy rate being 89.4%, demonstrating that the extracted energy signatures are effective rotation invariant features.