RBF network with two-stage supervised learning: an application

In the field of image classification, this paper compares a traditional RBF two-stage hybrid learning procedure with an RBF two-stage learning technique exploiting labeled data to adapt hidden unit parameters. RBF centers are determined by running a clustering algorithm separately on different training sets, where each set is associated with a different class. The ELBG neural network is used as clustering algorithm. Two different data sets have been considered. The first consists of real three Synthetic Aperture Radar (SAR) image tandem pairs from ERS1/ERS2 satellites. The second consists of Magnetic Resonance (MR) slices of a patient affected by multiple sclerosis. The results indicate that the supervised approach performs better than the traditional approach when the number of hidden unit is the same and seems more stable to changes in the number of hidden units.