This thesis introduces a method of texture segmentation, which is invariant with respect to orientation, scale and shift. The method is based on feature extraction by multi-channel Gabor filtering. The channels of the filter bank are organized in a polar-log scheme according to Fourier-Mellin approaches. The extracted features are classified with symmetric phase-only matched filtering. As these filters are optimal for the determination of peak location, orientation angle and scaling factor can be determined very precisely. Furthermore, the segmentation is insensitive with respect to noise. The thesis provides design criteria for filter banks. The segmentation algorithm automatically generates filter banks according to these criteria. The segmentation scheme consists of highly parallel and regular structures and is numerically efficient. A further speed-up is possible by employing sliding-window Fourier transform for Gabor filtering. The proposed segmentation scheme is suitable for the error detection in wood, cotton, textile or paper producing industries. It is suitable for texture-based object recognition and for aerial image analysis. The thesis shows segmentation results on textile textures, honing textures, Brodatz textures and artificial textures.