Stochastic modeling and handwritten letter recognition.

Summary Among the numerous handwritten letter recognition methods proposed in the last few years, a very limited number of them use or propose a modelization able to simulate the handwriting. This rarity is due to the difficulty to estimate, from a letter image, the trace of the strokes that compose it, because of many crossings and overlaps, trace from which an efficient modeling is possible. In this work, an algorithm for estimating the trajectories of features is described, based on a statistical study and a modeling of the crossings of features. Each letter is then characterized by a graph of curves. A stochastic modeling of this graph, using a mixture of Gaussian laws, is estimated by an em-type algorithm, on a large standard database. A recognition algorithm, based on this modeling, is described. For each letter image and for each letter class, an estimator of the curve graph is computed. A pre-estimation of this graph, using a regression on global characteristics of the letters, whose parameters have been estimated from simulation of letters, allows to obtain this estimator quickly. The identification of the class of the letter is obtained by a test on the likelihood of the graphs of curves. The results obtained allow to consider an extension of the model to handwritten words.