Practical Application of the Data Preprocessing Method for Kohonen Neural Networks in Pattern Recognition Tasks

摘要

Self-Organizing Map (SOM) is a very effective solution for solving pattern recognition problems. However, some ambiguities appear during learning process with the existence of linear patterns in the learning data, in this case, the learning process lasts for a long time and the network produces irrelevant results. The work provides the resolution of the detected problem and the application of the SOM for the pattern recognition. To achieve our objective and minimize the learning time, a SOM improved model has been developed. This model uses a special block able to filter the input data and reduce the size of the learning multitude. The presented experimental test results in this work show that the improved model exceeds the standard model in terms of the recognition results accuracy and the learning time. The results obtained in this work encouraged us to think about using the improved model to develop a smart approach (SmartMaps) of Geographic Information Systems (GIS).