Gabor filter design for unsupervised texture segmentation

摘要

This paper proposes a new Gabor filter design method for unsupervised texture segmentation of the textured image. In this paper, the texture characteristics, that are the most important factor in filter design stage, can be found by pre-clustering small blocks of an input image. The purposed system uses these results as the representation of textures for designing the filter. The segmentation result can be found by post-clustering the filter outputs. This met hod has no need for a priori of the texture properties. In the proposed method, the multi-channel filtering theory's orientation and frequency bandwidths setting reduce the number of free parameters which should be manually adjusted in the designing stage of the Gabor filters.