The Classification Step

The classification step is based on the use of artificial neural networks, or, in more disctinct terms, of multilayer perceptrons (MLPs). These ones realize a set of discriminant functions, that associate a score to each possible class. These scores may be regarded as being representative of the probability of each class to be the one of the character presented to the system.

Two kinds of multilayer perceptrons are available to obtain these values. Each of them is associated with a given kind of features, provided by the previous step of the recognition system. The Normalized Contour Analysis (NCA) method has appeared to be more efficient, and is the one that is applied with priority. Moreover, as the NCA method and the Averaged Pixel (AP) method are applied in combination, the recognition rate is significantly improved.

For instance, the misclassification error rate on handwritten digits has so been reduced of more than 40%, in relation to the one reached thanks to the use of the best of the individual MLPs. Moreover, so as to minimize the overall recognition time while retaining a high recognition rate, the use of the AP method is restricted and depends on the classification results obtained by applying first the NCA method only (figure 1).

Figure 1 - Cascade Combination of Multilayer Perceptrons.

In order to optimize the whole recognition process, several combination methods of multilayer perceptrons have been investigated. Depending on the practical application (digits recognition, uppercase letters recognition, or symbol recognition) several small multilayer perceptrons may be used in combination, in cascade or by substitution. For handwritten digits, a recognition system that is close to 50% faster than the most efficient of the individual MLPs, while retaining the recognition rate at its highest value, has so been developped.

The figure 2 shows samples of characters that were not correctly classified by using the best of the individual MLPs, but that are now correctly recognized, thanks to the use of several multilayer perceptrons in combination.

Figure 2 - Samples of digits that were not correctly classified by using the best of the individual MLPs, but that are now correctly recognized, thanks to the use of several multilayer perceptrons in combination. The correct classes are (from left to right, and from top to bottom): 1, 1, 2,3 3, 5, 6, 8, 9, 9.

Back to the Technologies page.

Back to the Handwritten Character Recognition Homepage.

send coments to gosselin@tcts.fpms.ac.be