Performance Comparison between Back-Propagation Learning and Kohonen Self-Organizing Neural Networks Algorithm in Terms of Pattern Recognition

Abstract

Pattern recognition using back-propagation learning and Kohonen self-organizing neural network algorithms has been developed and measured various performance based on different criteria and environment of the pattern. These pattern recognition systems have taken the object image as input. In image pre-processing stage, scaling and clipping process has been applied from the background image to avoid unnecessary portion of the object image. Feature extraction has been performed after applying filtering and edge-detection method. The extracted feature has been used as the input of the back-propagation learning neural networks (BPN) algorithm and Kohonen self-organizing mapping (SOM) algorithm. Networks have been trained to create the knowledgebase from the input features. Finally, these learned templates have been used for testing purpose. The difference between two training procedure is that the learned weights and thresholds have been updated to calculate the output for BPN and feature mapping technique in output grid has been used for Kohonen network. Finally, the performances of both algorithms have been measured and compared the learning and recognition performance on the various selected criteria.