Object recognition using self-organizing neural networks

An object recognition system using self-organizing neural networks is developed in this thesis. The basic structure of a single self-organizing neural network module consists of two layers: the preprocessing and classifier layers. The first layer maps a digitized grey level image into a two dimensional array of nodes which contains the extracted features of the original image. This array is called a self-organizing feature map, and its interconnections with the input are reinforced during training when input images are presented to the neural network without human supervision. The second layer functions as a lookup table and learns the extracted features of the self-organizing map for classification. The performance of the network has been evaluated with respect to the ability of recognizing objects using simulated data. The simulations include intensity sensitivity, size sensitivity, noise sensitivity, and input shape degradation tolerance. At the end of the thesis, a hierarchical structure which consists of several modules of self-organizing neural networks is proposed to enhance the efficiency and the recognition ability of the system.