HIERARCHICAL TEXTURE SEGMENTATION USING DICTIONARIES.pdfVIP

HIERARCHICAL TEXTURE SEGMENTATIONUSING DICTIONARIESP. Bajcsy and N. AhujaUniversity of Illinois at Urbana-ChampaignBeckman Institute405 N. Mathews, Urbana, IL 61801AbstractWe present a new hierarchical texture segmenta-tion method that partitions an image into textured re-gions. A textured region is viewed as a set of uni-formly distributed primitives. A primitive is a regionwith constant gray values. Gray values within a prim-itive can be corrupted by noise. Any noisy primi-tive contains gray values from a -wide interval (-homogeneous primitive). The noisy primitive is de-scribed by the mean of interior gray values. A texturedregion with noise is characterized by a set of gray valuemeans (texture dictionary) derived from noisy primi-tives. Every pixel (sample point) and its neighborhoodgive rise to an estimate of texture dictionary. Com-ponents of the estimated dictionary at a pixel char-acterize noisy primitives of a textured region grownfrom the pixel. Co-occurrence of noisy primitives fromthis grown region are calculated. Final segmentation isobtained by grouping pixels with identical dictionariesand co-occurrences created at each pixel. Homogene-ity degree  of noisy primitives provides a frameworkfor multiscale analysis. Computational eciency androbustness of the proposed method are related. Experi-ments are reported for synthetic and real textures fromBrodatz album and real scenes.1 IntroductionThe goal of any image segmentation is to partitionan image (grid of samples) into connected subsets ofsamples, denoted as regions, each having a uniformtexture. Textures have no universal model. A varietyof texture models have been derived from: (1) per-ceptual studies [9, 8, 6] (mimicking humans), (2) spe-cic two-dimensional tasks [13], such as, automatedsurface inspection (textile, paint), medical image pro-cessing (semi-automated search for tissues, tumors),(3) texture gradient analyses [3, 2] (projective distor-tion problem) and (4) texture imitation [10, 5] (r