Improvement of On-line Recognition Systems Using a RBF-Neural Network Based Writer Adaptation Module

In this paper we designed an adaptation module (AM) with the objective to increase the performance of a recognition system for a new user or new writing style. The developed adaptation module is added after the recognition system, and its role is to examine the output of the independent system and produce a more correct output vector close to the desired response of the user. To achieve this end, we conceive an adaptation module based on Radial Basis Function Neural Network (RBF-NN) which is built using an incremental training algorithm. Two adaptation strategies are applied for adaptation module training: increase the number of new hidden units and adjust the parameters of the nearest unit (weights and location of center) using the standard descent gradient. This new architecture is evaluated by the adaptation of two recognition systems, one for digit recognition and one for alphanumeric character recognition. The results, reported according to the cumulative error, show that the adaptation module (AM) leads to decreasing the classification error and is capable of fast adaptation to the users handwriting. Moreover, results are compared with those carried out using the weights updating strategy of the nearest center apart from the addition of new units. In fact, the adaptation module decreases an average of 50% the error rate with standard recognition systems.