On the Choice of Training Set, Architectures and Combination Rule of Multiple MLP Classifiers for Multiresolution Recognition of Handwritten Characters
Abstract
In a number of recognition approaches for recognition of handwritten characters, MLP networks have been used successfully. However, to the best of our knowledge, there does not exist a detailed study about the selection of network architecture and training strategy. In this article, we considered wavelet transform-based multiresolution features along with multiple MLP classifiers. Our study shows that huge number of nodes in the hidden layer provides high recognition performance. Secondly, for the training, it is observed that a dynamic selection strategy of the training samples can avoid unnecessary expensive computation and still provide perhaps better generalization. We also studied a number of approaches for the combination of the multiple MLP classifiers and we obtained recognition accuracy comparable to the state-of-the-art technologies. This approach has been tested on MNIST database for handwritten English digits and obtained approximately 99% recognition accuracy on the test set.