: Bresch, M.; Hosticka, B.J.; Schrey, O.

Abstract
A method for rotation and scale invariant texture segmentation is proposed, which can be also employed for object recognition based on pattern analysis in noisy images. The segmentation scheme is based on a supervised rotation and scale invariant texture recognition using multichannel polar logarithmic Gabor filters for feature extraction. The polar logarithmic arrangement works like a Fourier-Mellin descriptor providing orientation and scale invariance. The classification of the features is carried out by symmetric phase-only matched filtering. The classification accuracy is about 90% at arbitrary rotation angle and for scale factors between 0.25 and 4.0. Rotation angle and scale factor can be determined with high precision by the classification scheme. Prior to the segmentation a normalization scheme as preprocessing step is used to reduce illumination gradients, which is also able to treat illumination edges like shades.