It is found that a small subset of the investigated texture features is highly significant for all experiments. Using this feature subset in conjunction with multi-layer, nonlinear and low complexity feed-forward network architectures classification accuracies in the order of 99% are obtained. |