LIP6 2000/031
- Thesis
Modélisation de séquence par techniques adaptatives : prévision de décharges de batterie et extraction de contours dans des images médicales - O. Gérard
- 241 pages - 06/29/1999- document en - http://www.lip6.fr/lip6/reports/2000/lip6.2000.031.ps.gz - 5,588 Ko
- Contact : gerard (at) nulllep-philips.fr
- Ancien Thème : APA
- Keywords : Neural network, adaptive method, discharge prediction, rechargeable batteries, contour detection, medical images
- Publisher : Valerie.Mangin (at) nulllip6.fr
The aim of this thesis is to propose and develop a general method for analyzing two kinds of time series problems. This general approach is based on a modelization of the observed phenomenon and on statistical learning of the model parameters and their evolution. For these two problems, hierarchical hybrid systems have been developed using the universal approximation capability of the neural networks and integrating prior knowledge about the problem.
The first application is the prediction of the behavior of a dynamic system evolving under the influence of changes in its environment. More precisely, the goal of our proposed hierarchical and evolving system was to predict the end of the discharge of batteries powering a portable system. This proposed system consists in two neural networks hierarchically organized. The first network is a simple model of a discharge curve while the second is responsible for the adaptation to the context by the estimation the weights of the first network. An incremental version further adapts on-line the system to take into account individual behavior of the batteries. The results obtained are rather good with an average error of about 6 minutes for an event that may occur within 10 hours.
The second application deals with the automatic extraction of a contour in medical images and more precisely with the detection of the left ventricle contours in angiocardiographic images. The precise determination of this contour is the basis of several measures required for diagnosing cardio-vascular diseases. The proposed system makes use of prior high-level knowledge to direct the search toward the most probable contour. This search is based on a hybrid neural network - hidden Markov model system. The promising results demonstrate the validity of the developed approach.
The first application is the prediction of the behavior of a dynamic system evolving under the influence of changes in its environment. More precisely, the goal of our proposed hierarchical and evolving system was to predict the end of the discharge of batteries powering a portable system. This proposed system consists in two neural networks hierarchically organized. The first network is a simple model of a discharge curve while the second is responsible for the adaptation to the context by the estimation the weights of the first network. An incremental version further adapts on-line the system to take into account individual behavior of the batteries. The results obtained are rather good with an average error of about 6 minutes for an event that may occur within 10 hours.
The second application deals with the automatic extraction of a contour in medical images and more precisely with the detection of the left ventricle contours in angiocardiographic images. The precise determination of this contour is the basis of several measures required for diagnosing cardio-vascular diseases. The proposed system makes use of prior high-level knowledge to direct the search toward the most probable contour. This search is based on a hybrid neural network - hidden Markov model system. The promising results demonstrate the validity of the developed approach.