Outcomes of Neural and Rule-based Classifiers as Criterions in Bi-objective Evolutionary Optimization of Feature Space in Pattern Recognition

Neural networks are widely used as classifiers in many pattern recognition problems because of good generalization abilities, what is a crucial issue in any practical application. However, vast majority of neural network architectures demands a huge computational effort for the training process, what in turn limits such solutions from application in one important domain of pattern recognition, which is the optimization of feature extractors. The optimization requires iterative computation of the objective function, and therefore such computation should be univocal and computationally effective. The paper presents how these postulates are satisfied in a special neural architecture called probabilistic neural network, which can therefore be effectively used for calculation of criterion in an evolutionary optimization of the feature space. As the experimental verification of proposed methodology, the optimization of Fraunhofer diffraction based pattern recognition system is presented, and compared with alternative solution, i.e. application of the rule-based classifier outcomes in the same role. The optimized system is a class of hybrid, fast, opto-electronic image recognizers, and the paper presents the use of it in the recognition of three different domains of images: recognition of the type of the vehicles, recognition of the type of the road obstacle in infrared wavelength, and the recognition of the class of the subsurface stress in the optical fiber. The experimentally obtained results confirm, that probabilistic neural network based main criterion used in a bi-criterion evolutionary optimization, outperforms the rough set based criterion in the mentioned systems. Key-Words: neural networks, rough sets, bi-criterion optimization, evolutionary optimization, holographic ring-wedge detectors, image processing system, pattern recognition, infrared imaging, hybrid systems