New cost function for backpropagation neural networks with application to SAR imagery classification
Hossam M. Osman, Steven D. Blostein · Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE · 1999
This paper proposes the minimization of a new cost function while training backpropagation (BP) neural networks to solve pattern classification problems. The new cost function is referred to as the gain-weighted normalized-target mean-square error (GWNTMSE). The paper proves that the minimization of the GWNTMSE is optimal in the sense of yielding network classifier with minimum variance from the optimal Bayes classifier in the limit of an asymptotically large number of statistically independent training patterns. Experimental results are presented. The application selected is the classification of ship targets in airborne synthetic aperture radar (SAR) imagery. The number of ship classes is 8. They represent 2 destroyers, 2 cruisers, 2 aircraft carries, a frigate, and a support ship. The obtained results indicate that BP classifiers trained by minimizing the GWNTMSE consistently outperform those trained by minimizing the standard MSE.