Generating rotation-invariant texture features by randomization of operator orientation

Rotation-invariant texture features are generated by randomizing the orientation of the underlying texture operators. The approach is applied to texture features based on local binary patterns as well as to sum and difference histograms. Results are given for a difficult classification problem of 15 different Brodatz textures and 7 rotation angles. Due to randomamization, the error rate becomes independet of the texture orientation. Moreover, the classification of periodic textures is enhanced significantly.