Rough Neural Networks Architecture For Improving Generalization In Pattern Recognition

Neural networks are found to be attractive trainable machines for pattern recognition. The capability of these models to accommodate wide variety and variability of conditions, and the ability to imitate brain functions, make them popular research area. This research focuses on developing hybrid...

Full description

Saved in:
Bibliographic Details
Main Author: Ali Adlan, Hanan Hassan
Format: Thesis
Language:English
English
Published: 2004
Online Access:http://psasir.upm.edu.my/id/eprint/5116/1/FK_2004_91.pdf
http://psasir.upm.edu.my/id/eprint/5116/
Tags: Add Tag
No Tags, Be the first to tag this record!
Institution: Universiti Putra Malaysia
Language: English
English
id my.upm.eprints.5116
record_format eprints
spelling my.upm.eprints.51162013-05-27T07:20:30Z http://psasir.upm.edu.my/id/eprint/5116/ Rough Neural Networks Architecture For Improving Generalization In Pattern Recognition Ali Adlan, Hanan Hassan Neural networks are found to be attractive trainable machines for pattern recognition. The capability of these models to accommodate wide variety and variability of conditions, and the ability to imitate brain functions, make them popular research area. This research focuses on developing hybrid rough neural networks. These novel approaches are assumed to provide superior performance with respect to detection and automatic target recognition.In this thesis, hybrid architectures of rough set theory and neural networks have been investigated, developed, and implemented. The first hybrid approach provides novel neural network referred to as Rough Shared weight Neural Networks (RSNN). It uses the concept of approximation based on rough neurons to feature extraction, and experiences the methodology of weight sharing. The network stages are a feature extraction network, and a classification network. The extraction network is composed of rough neurons that accounts for the upper and lower approximations and embeds a membership function to replace ordinary activation functions. The neural network learns the rough set’s upper and lower approximations as feature extractors simultaneously with classification. The RSNN implements a novel approximation transform. The basic design for the network is provided together with the learning rules. The architecture provides a novel method to pattern recognition and is expected to be robust to any pattern recognition problem. The second hybrid approach is a two stand alone subsystems, referred to as Rough Neural Networks (RNN). The extraction network extracts detectors that represent pattern’s classes to be supplied to the classification network. It works as a filter for original distilled features based on equivalence relations and rough set reduction, while the second is responsible for classification of the outputs from the first system. The two approaches were applied to image pattern recognition problems. The RSNN was applied to automatic target recognition problem. The data is Synthetic Aperture Radar (SAR) image scenes of tanks, and background. The RSNN provides a novel methodology for designing nonlinear filters without prior knowledge of the problem domain. The RNN was used to detect patterns present in satellite image. A novel feature extraction algorithm was developed to extract the feature vectors. The algorithm enhances the recognition ability of the system compared to manual extraction and labeling of pattern classes. The performance of the rough backpropagation network is improved compared to backpropagation of the same architecture. The network has been designed to produce detection plane for the desired pattern. The hybrid approaches developed in this thesis provide novel techniques to recognition static and dynamic representation of patterns. In both domains the rough set theory improved generalization of the neural networks paradigms. The methodologies are theoretically robust to any pattern recognition problem, and are proved practically for image environments. 2004 Thesis NonPeerReviewed application/pdf en http://psasir.upm.edu.my/id/eprint/5116/1/FK_2004_91.pdf Ali Adlan, Hanan Hassan (2004) Rough Neural Networks Architecture For Improving Generalization In Pattern Recognition. PhD thesis, Universiti Putra Malaysia. English
institution Universiti Putra Malaysia
building UPM Library
collection Institutional Repository
continent Asia
country Malaysia
content_provider Universiti Putra Malaysia
content_source UPM Institutional Repository
url_provider http://psasir.upm.edu.my/
language English
English
description Neural networks are found to be attractive trainable machines for pattern recognition. The capability of these models to accommodate wide variety and variability of conditions, and the ability to imitate brain functions, make them popular research area. This research focuses on developing hybrid rough neural networks. These novel approaches are assumed to provide superior performance with respect to detection and automatic target recognition.In this thesis, hybrid architectures of rough set theory and neural networks have been investigated, developed, and implemented. The first hybrid approach provides novel neural network referred to as Rough Shared weight Neural Networks (RSNN). It uses the concept of approximation based on rough neurons to feature extraction, and experiences the methodology of weight sharing. The network stages are a feature extraction network, and a classification network. The extraction network is composed of rough neurons that accounts for the upper and lower approximations and embeds a membership function to replace ordinary activation functions. The neural network learns the rough set’s upper and lower approximations as feature extractors simultaneously with classification. The RSNN implements a novel approximation transform. The basic design for the network is provided together with the learning rules. The architecture provides a novel method to pattern recognition and is expected to be robust to any pattern recognition problem. The second hybrid approach is a two stand alone subsystems, referred to as Rough Neural Networks (RNN). The extraction network extracts detectors that represent pattern’s classes to be supplied to the classification network. It works as a filter for original distilled features based on equivalence relations and rough set reduction, while the second is responsible for classification of the outputs from the first system. The two approaches were applied to image pattern recognition problems. The RSNN was applied to automatic target recognition problem. The data is Synthetic Aperture Radar (SAR) image scenes of tanks, and background. The RSNN provides a novel methodology for designing nonlinear filters without prior knowledge of the problem domain. The RNN was used to detect patterns present in satellite image. A novel feature extraction algorithm was developed to extract the feature vectors. The algorithm enhances the recognition ability of the system compared to manual extraction and labeling of pattern classes. The performance of the rough backpropagation network is improved compared to backpropagation of the same architecture. The network has been designed to produce detection plane for the desired pattern. The hybrid approaches developed in this thesis provide novel techniques to recognition static and dynamic representation of patterns. In both domains the rough set theory improved generalization of the neural networks paradigms. The methodologies are theoretically robust to any pattern recognition problem, and are proved practically for image environments.
format Thesis
author Ali Adlan, Hanan Hassan
spellingShingle Ali Adlan, Hanan Hassan
Rough Neural Networks Architecture For Improving Generalization In Pattern Recognition
author_facet Ali Adlan, Hanan Hassan
author_sort Ali Adlan, Hanan Hassan
title Rough Neural Networks Architecture For Improving Generalization In Pattern Recognition
title_short Rough Neural Networks Architecture For Improving Generalization In Pattern Recognition
title_full Rough Neural Networks Architecture For Improving Generalization In Pattern Recognition
title_fullStr Rough Neural Networks Architecture For Improving Generalization In Pattern Recognition
title_full_unstemmed Rough Neural Networks Architecture For Improving Generalization In Pattern Recognition
title_sort rough neural networks architecture for improving generalization in pattern recognition
publishDate 2004
url http://psasir.upm.edu.my/id/eprint/5116/1/FK_2004_91.pdf
http://psasir.upm.edu.my/id/eprint/5116/
_version_ 1643823095243341824