Design and evaluation of neural networks for an embedded application

Abstract

In this study we compare different training strategies and configurations of Artificial Neural Networks (ANNs) designed for a real application, a mobile reader device for blind people. They are used to recognize characters found in input images; since the application has a real-time behaviour, we have improved the already implemented ANN subsystem rather than looking for new and more complex solutions. The main idea has been to setup a pool of configurations and then to select the best one through a rigorous test. All the results have been computed on both the development platform and the target one, considering implementation issues. This method has led to improve the performances of the classifier maintaining the same complexity as the starting one. © 2010 Springer-Verlag.