Search results for key=Squ1996 : 1 match found.

In this thesis, the notion of Model-Based Neural Networks is introduced. Model-Based Neural Networks, whilst retaining the essential structure and advantages of traditional neural networks, enable explicit modeling of the target system. This can result in dramatically improved generalization of classification performance to patterns not present in the training data, and representation in considerably lower-dimensional state spaces.

An important problem in many areas of computer vision is invariant pattern recognition. This is the chosen domain for demonstrating the efficacy of the Model-Based Neural Network approach. Model-Based Neural Networks can be constructed which have responses which are invariant to specified transformations of the input data. Such networks can be trained with much smaller training sets than are required by traditional networks, since it is no longer necessary to provide examples of transformed versions of the input prototypes. This, coupled with the reduction in the dimensionality of the parameter space, means that training such networks is often much less computationally-expensive than the traditional alternative.

To situate this work, a review of existing general techniques for invariant pattern recognition is presented in Chapter 2, and of previous neural network-based approaches in Chapter 3. Chapter 4 presents a variety of different forms of Model-Based Neural Network, and demonstrates their utility on a range of invariant pattern recognition problems, both real and synthetic. Included is a comparison with an earlier study, which reveals the great improvements possible with Model-Based Neural Networks.

Chapter 5 introduces a new invariant feature of two-dimensional contours, the Invariance Signature. It is shown in Chapter 6 that a Model-Based Neural Network can be constructed which calculates this multi-dimensional feature, and classifies patterns on this basis. Chapter 7 reports experimental results demonstrating that such Invariance Signature-based MBNNs can be employed successfully for shift-, rotation- and scale-invariant optical character recognition.