Handwritten character recognition using the "Championship Algorithm"

In the last few years many methods for OCR (Optical Character Recognition) employing Artificial Neural Networks (AN N) have been proposed [5]. An AN N can be used either for the classification step or for the complete recognition system. The input to the AN N may be represented by some "features" of the character, similar to those used in statistical and syntactical OCR systems [3]. The input is commonly sizenormalized and the outputs are exclusively coded. Many researchers have experimented fully-connected multilayered networks trained with Backpropagation. One very serious problem with these classifiers is the high number of free parameters due to the high number of outputs (classes), inputs and hidden neurons required for the convergence of Backpropagation [6]. There are several attempts to face this problem. One remarlmble solution seems to be that of imposing some constraints on the network architecture in order to specify priori knowledge on the task ([1], [2]). Another one is that of using the "divide et impera" paradigm and consequent1y that of exploring modular architectures [7]. In order to face these scaling-up problems, in this paper .we propose a modular technique referred to as the "championship algorithm". It is based on autoassociators that are AN N with as many outputs as inputs, that are trained to reply the input pattern to the output units. To use autoassociators for character classification we can construct and train in the learning phase an autoassociator for each class. In the recognition phase we feed the test character to each autoassociator and evaluate the mean square error between the inputs and the outputs. Lower error values allows us to identify classes that are more similar to the test character. Theautoassociator with the lowest error is likely to corresponds with the right class. However a similar technique has not high discrimination capabilities.