Invariant texture segmentation using multi-channel approaches

A method of rotation and scale invariant texture segmentation is proposed, which can also be employed for object recognition based on pattern analysis in noisy images. The method has low sensitivity to illumination and low noise sensitivity. With this method, orientation angle and scaling factor can be determined as well. The texture segmentation proposed in this paper works on dividing an image into a set of small regions and classifying each region. The proposed texture segmentation method is based on a supervised rotation, scale, and shift invariant texture recognition scheme. In this scheme, features of data base textures and features of a texture to be recognized are extracted with a common definite polar-log Gabor filter bank and classified with symmetric phase-only matched filtering. For reducing the numerical costs of the segmentation, a method for selecting pixels/small regions to be classified is introduced. The numerical efficiency of spatial filtering is compared with filte ring results. In this paper, three methods reducing illumination gradients are proposed, which can be applied as proprocessing steps for texture segmentation.