ABSTRACT
This section describes the use of a hierarchy of neural networks, in association with nonneural techniques, for the accurate segmentation of natural textured images. The segmentation is unsupervised as only the total number of different textures is prespecified for a particular input image and no exemplars of the expected textures are provided. The system uses a model estimator, based on a Markov random field model, for texture parameter extraction. Two unsupervised neural networks, a Kohonen self-organizing map and a local-voting network, are employed to produce a ‘clean’ segmented image. A further refinement to improve on the detailed positioning and form of the texture boundaries is provided by a boundary relaxation algorithm. Typical results and details of operational parameters are provided.
