Ranking ICA bases by associative memory recalls of training texture samples

We wish to generalize the covariance matrix approach (PCA) by the statistical Independent Component Analyses (ICA), which have been implemented by Bell-Sejnowski efficiently using ANN methodology. The gain of the statistics is the los of the geometry. In this research, we preserve the texture geometry with a so-called local ICA, in order to extract separately independent features from each class of natural textures. To avoid the curse of the dimensionality due to the local ICA, we furthermore use the divide-and-conquer strategy. A single ICA basis vector is chosen from each texture class, based on the maximum associative recalls from the class training set. Subsequently, another ICA basis is chosen, if necessary, to minimize the false alarm rate, namely the spread of confusion matrix. For the visible remote sensing application, we have designed such an optimum classifier of all natural scene textures with a minimum spread of the confusion matrix.