Use this URL to cite or link to this record in EThOS: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.604678
Title: Alternative Bayesian techniques for model selection, classification, and parameter estimation in signal and image processing
Author: Hsieh, M. C.-M.
Awarding Body: University of Cambridge
Current Institution: University of Cambridge
Date of Award: 2000
Availability of Full Text:
Full text unavailable from EThOS.
Please contact the current institution’s library for further details.
Abstract:
This thesis intends to address some key aspects of the implementation of Bayesian analysis for classification, model selection and parameter estimation, and proposes three enhancements or alternative approaches to commonly employed techniques. 1. An extension of the General Linear Model allows prior parameter information to be included whilst retaining the analytic and accessible form of the model evidence and posterior distribution. Channel estimation in non-stationary noise and retrospective excitation changepoint detection are used as illustrations of the extended model's applicability. 2. A polynomial approximation to the likelihood function allows marginalised estimates of model parameters to be obtained in the form of a Volterra series. The series can be applied directly to the observed data vector in an iterative fashion, to converge upon a set of parameter maximum a-posteriori (MAP) estimates with low computational cost. An example implementation for optical character recognition (OCR) of handwritten characters illustrates the behaviour and utility of the estimator. 3. An optimisation algorithm is proposed based on recursive ordering of the model parameter marginals for model selection. The marginals are computed using Markov chain based simulated annealing, which in itself is an effective optimisation algorithm. Results for model candidates based on highly correlated linear basis functions show that the recursive parameter ordering algorithm enhances the performance of the simulated annealer over that of simulated tempering for individual sampling runs. Optimisation based upon this algorithm can work comparatively more efficiently when the observed data is approximately locally rather than globally correlated. This advantage is demonstrated in an overlapping object recognition example where the ability of the Volterra parameter estimator to handle obscured data is also utilised.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID: uk.bl.ethos.604678  DOI: Not available
Share: