A Method for the Improvement of the Behavior of Bidirectional Associative Memories as Pattern Classifiers

摘要

We present a form of attaining success levels of up to 100% in character classification by the appropriate use of thresholds in the activity functions of the neurons making up the two-layer network with which bidirectional associative memories are implemented, together with a systematic method for generating the weight matrix. The system that is constructed includes a geometrical pre-processing stage that eliminates distortions, thereby improving the results. As a final characteristic, the functioning of the system presents a high level of immunity to noise or deformations. The system was evaluated using the two popular databases NIST#19 and UCI. There was found to be no misclassification in any case, whether under conditions of heavy contamination from noise or distortion of the image to be classified.