Hierarchical texture segmentation

Abstract

We present a new hierarchical texture segmentation method that partitions an image into textured regions. A textured region is viewed as a set of uniformly distributed primitives. A primitive is a region with constant gray values. Gray values within a primitive can be corrupted by noise. Any noisy primitive contains gray values from a δ-wide interval (δ-homogeneous primitive. The noisy primitive is described by the sample mean of interior gray values. A textured region with noise is characterized by a set of gray value sample means (texture vector) derived from noisy primitives. Every pixel (sample point) and its neighborhood give rise to an estimate of texture vector. Components of the estimated vector at a pixel characterize noisy primitives of a textured region grown from the pixel. Co-occurrence of noisy primitives from this grown region are calculated. Final segmentation is obtained by grouping pixels with identical estimates of texture vectors and co-occurrences, created at each pixel. Homogeneity degree δ of noisy primitives provides a basis for multiscale analysis. Computational efficiency and robustness of the proposed method are related. Experiments are reported for synthetic textures as well as real textures from Brodatz album and real gray scale and color images.