Data reduction via auto-associative neural networks

Image analysis is a very complex process; many of the relationships are difficult to categorize, much less to program into a computer. The selection of features is the most challenging problem of image analysis, process discovery or sensor fusion. The features must be a data representation that will discriminate the information of interest from the rest of the image. A neural network can be a tool for rapid processing of data. Auto-associative neural networks (AANNs) are a form of self-organizing maps which can be used to reduce the dimension of the input data in a self-organizing fashion. Dimension reduction is closely related to feature extraction. Features are those datum that efficiently capture the information contained in the entire data set. The data set, has a “superficial” dimensionality of n, and the reduced space of the features that contain all the information about the data has an “intrinsic” dimension of m, where n>m. With this property, an AANN can be used to reduce an n-dimensional space to something more intrinsic to the actual data