Skip to Main Content
The article outlines a methodology to automatically select a neural-based pattern classifier. A set of neural-based specialized pattern recognizers is generated, trained and successively it is automatically chosen, among them, one which has the "best" generalization capabilities according to a quality index, that does not require the use of any test set. Furthermore, it is illustrated the architecture of a basic element, based on the EaNet neural classifier, of a more complex framework that will be designed for concurrent pattern recognition in networked repositories of patterns. The effectiveness of the proposed approach has been tested as an example on the "NIST Special database 19" of handwritten characters images and it has also been verified using the traditional technique of the test set. For completeness, the methodology has been also tested using a traditional neural feed-forward classifier using sigmoids as activation function of its units belonging to the hidden layer. Experimental results show good performance of the proposed methodology.