Radar clutter classification based on noise models and neural networks

The problem of the classification of clutter textures in coherent images, e.g. synthetic aperture radar (SAR), is discussed. The first stage in such a process is to segment the observed texture into regions of differing texture properties. In order to select the segmentation regions one must exploit some information about the properties of the texture. One method for encapsulating this information is in the form of a model which is known a priori. Another approach is to train a segmentation method by large numbers of examples of the types of texture encountered. This latter approach underlies the use of noncommittal neural networks. A comparison of these approaches reveals the extent to which a neural network is capable of extracting all the information contained within the texture which is automatically exploited in the model-based approach.<>