Abstract
This paper describes a method for lexical filtering to be employed mainly (but not exclusively) in cursive word recognition. This method operates at word level by selecting a subset of the original vocabulary based on a simple and robust characterization of the image of the unknown word. The features adopted for word representation are basically morphological (ascenders, descenders, loops) or structural (hypothesized number of letters). In this work we have added another feature to improve the efficiency of the lexical filter, namely the recognition of the initial letter. In order to classify the initial letter reliably, without segmentation or ad-hoc training, we applied an MLP-based character spotting method.
Hence, depending on wether intensity (nonlinear, easier to perform) or field (linear) measures are chosen, Newton [3] or Gauss-Sedel [4] methods can be used to solve the problem. Both are affected by inherent limitations. Convergency to spurious (local) minima for Newton’s method, and ill-conditioning for Gauss-Seidel method show up as the distance between the source and the measurement plane is increased.
During the last few years, a number of hopefully more effective tools for (global) minimization problems where many local minima do exist have been developed, including both stochastic (genetic algorithims [5], simulated annealing [6], Boltzmann machines [6] and deterministic (mean field neural networks [7] methods.
In this paper we adopt Vidyasagar’s [8] (nonlinear) mean field neural network, which can be regarded as mean-field version of Boltzmann machine, in view of its comparative speed, to solve the antenna array diagnostic problem. An asynchronous, siscrete implementation of the (originally continuous) Vidyasagar’s mean field nueral network is introduced, which allows to define a Lyapounov function, whereby the convergency to a stationary point can be proven. For a suitable choice of the neural threshold function, the equilibrium state is shown to be unique.
The performance of the algorithm is investigated for different measurement weight factors, and varying values of the radio d/D, dbeing the distance between the source and the measurement plane, and D the array diameter. It is shown that a judicious choice of the weights leads in a few steps to the correct solution with probability one as d/D → 0.
This is a preview of subscription content, log in via an institution.
Buying options
Tax calculation will be finalised at checkout
Purchases are for personal use only
Learn about institutional subscriptionsPreview
Unable to display preview. Download preview PDF.
References
Leroy A. Progressive Lexicon Reduction for On-Line Handwriting. Progress in Handwriting Recognition, pp. 399–406, World Scientific, 1997
Madhvanath S. and Govindaraju V. Holistic Lexicon Reduction. IWFHR, pp. 71–81, 1993
Bellaby GJ., Evett LJ. and Powalka RK. Coping with Ambiguity and Error in Cursive Script Recognition. Progress in Handwriting Recognition, pp. 337–344, World Scientific, 1997
Hall PAV. and Dowling GR. Approximate string matching. Computing Surveys 12(4), pp. 381–402, 1980
Colla AM. and Pedrazzi P. Single and Coupled Neural Handprinted Character Classifiers. In: M. Marinaro and P.G. Morasso (Ed.s), ICANN 94 — Proc. of the Int. Conf. on ARTIFICIAL NEURAL NETWORKS. Sorrento (Italy), vol. II, pp. 969–972, Springer-Verlag, 1994
Camastra F., Cepollina E. and Colla AM. Word Recognition by MLP-based Character Spotting and Dynamic Programming. In M. Marinaro and R. Tagliaferri (Ed.s), Neural Nets. WIRN VIETRI — 97, Proc. 9th Italian Workshop on Neural Nets, Vietri s.M., Italy, 22–24 May 1997, pp. 183–188, Springer-Verlag, 1998
Lorenzon A. Sviluppo di un filtro lessicale per riconoscitori di parole. Tesi di Laurea in Informatica (M.Sci.), Università di Genova, Dec. 1997
Colla AM. and Lorenzon A. A simple lexical filter for two-step recognition of cursive words (in preparation)
Author information
Authors and Affiliations
Editor information
Editors and Affiliations
Rights and permissions
Copyright information
© 1999 Springer-Verlag London Limited
About this paper
Cite this paper
Colla, A.M., Lorenzon, A. (1999). Initial Letter Spotting as a Complementary Feature for Lexical Filtering of Cursive Words. In: Marinaro, M., Tagliaferri, R. (eds) Neural Nets WIRN VIETRI-98. Perspectives in Neural Computing. Springer, London. https://doi.org/10.1007/978-1-4471-0811-5_18
Download citation
DOI: https://doi.org/10.1007/978-1-4471-0811-5_18
Publisher Name: Springer, London
Print ISBN: 978-1-4471-1208-2
Online ISBN: 978-1-4471-0811-5
eBook Packages: Springer Book Archive