The Technologies We Use
The main structure of an off-line handwritten character recognition system has already been presented. They are three steps, which are here reminded, and more precisely described.
The first step consists in analysing the digitalized image provided by a scanning device, so as to isolate the characters one from each others. Two algorithms have especially been developped.
The second step is a discriminant feature extraction process, for which several methods have been deeply analysed. The two most efficient of these methods have been improved further, by using complementary techniques. This has led to two new methods which appears to be extremely efficient.
The third step is the classification step, by itself. It is based on the use of Multilayer Perceptrons (MLPs). The non-linearities that are included in these connectionist sytems, and the discriminant trainning phase that they are submitted to, leads these kinds of Artifical Neural Networks particularly suitable for classification tasks. Two categories of Multilayer Perceptrons have been trained, each of these one being associated with one of the feature vector provided by the previous step. So as to optimize the recognition step, several methods of combination of Multilayer Perceptrons have been set out. Recognition tests have then shown that cooperation of neural networks that possess disctinct types of features could reduce significantly the overall error rate. The actual recognition system consists of a cascade association of small MLPs, which allows minimization of the overall recognition time while retaining a high recognition rate.
Finally, a new character distorsion method, allowing to artificially create handwritten character images from real ones, has been developped. This method can be applied to increase the diversity of the database that is used for training the neural networks, which may leads to a significant improvement of their classification performance, without altering their computation time during the recognition phase.
The main part of the work has been the developpement of efficient feature extraction methods on the one hand, and the setting of Multilayer Perceptrons combination methods for classification on the other hand. The association of these techniques has allowed recognition rates of 96.7% and 99.8% to be reached, respectively for unconstrained handwritten uppercase letters and digits extracted from the NIST3 database. When they have been applied to handwritten digits of "European" type, whose aspects are very more diversified, these methods have offered a recognition rate of 95.3%. Relatively to the best of the individual multilayer perceptron, this represents a misclassification error rate that is 40% lower, as well as a recognition process that is 50% faster.
Back to the Handwritten Character Recognition Homepage.
Send your comments to firstname.lastname@example.org