Automatic neural generalized font identification
Neural methods are gaining a steady acceptance as powerful tools in a variety of pattern detection problems, OCR certainly being one of them. The concrete implementation of these neural OCR systems is of course a well guarded corporate secret, but in broad terms it can be said that in most of the cases, multilayer perceptrons (MLPs) are used. There are several reasons for the MLPs’ success. To begin with, they are based in well understood mathematical and statistical principles and there are efficient tools and methodologies for their training and evaluation. Furthermore they have good generalization properties.
KeywordsDiscrete Cosine Transform Radial Basis Function Network Inverse Discrete Cosine Transform Model Overfitting Good Generalization Property
Unable to display preview. Download preview PDF.
- A. Sierra, C. Santa Cruz, V. López, G. Fractman, J. Dorronsoro, C. Aguirre, J.M. Soto, A. Medina, R. López, “Neural networks in large scale bank effect recognition”, Procs. SNN 1995 Symposium on Neural Networks, B. Kappen, S. Gielen (eds), 393–396, Springer Verlag, 1995.Google Scholar
- L. Xu, M.I. Jordan, “On convergence properties of the EM algorithm for Gaussian mixtures”, Neural Computation 7 (1995), 129–151.Google Scholar