A Hybrid Neural Network Architecture for Texture Analysis in Digital Image Processing Applications

Abstract

A new hybrid neural network model capable of texture analysis in a digital image processing environment is presented in this thesis. This model is constructed from two different types of neural network, self-organisation and back-propagation. Along with a brief resume of digital image processing concepts, an introduction to neural networks is provided. This contains appropriate documentation of the neural networks and test evidence is also presented to highlight the relative strengths and weaknesses of both neural networks. The hybrid neural network is proposed from this evidence along with methods of training and operation. This is supported by practical examples of the system's operation with digital images. Through this process two modes of operation are explored, classification and segmentation of texture content within images.

Some common methods of texture analysis are also documented, with spatial grey level dependence matrices being chosen to act as a feature generator for classification by a back-propagation neural network, this provides a benchmark to assess the performance of the hybrid neural network. This takes the form of descriptive comparison, pictorial results, and mathematical analysis when using aerial survey images.

Other novel approaches using the hybrid neural network are presented with concluding comments outlining the findings presented within this thesis.