Coalescing Texture Descriptors

ARPA Image Understanding Workshop 1996 As an alternative to texture segmentation for the description of the images, we propose a method for coalescing descriptors of adjacent image patches with similar textural content into tight clusters. We achieve this result by extending the notion of edge-preserving smoothing and anisotropic di usion from gray-level and color images to vector-valued images that describe the textural content of each image patch. To this end, we rst compute raw texture descriptors, that is, vectors that describe local texture appearance. We then de ne texture contrast as the maximum derivative of the texture descriptor vector, and we de ne signi cant regions as those in which contrast is low. We propose an e cient implementation of edge-preserving smoothing and show that it is closely related to anisotropic di usion. Experiments on texture mosaics and real images show that clear texture edges are found even where intensity edges are weak or poorly de ned, and that signi cant texture regions yield tight clusters of texture descriptors.