Refinements in iterative blind deconvolution (1995)
Type of ContentTheses / Dissertations
Thesis DisciplineElectrical Engineering
Degree NameDoctor of Philosophy
PublisherUniversity of Canterbury
AuthorsJiang, Hongshow all
The thesis presents a number of results regarding blind deconvolution of a single contaminated blurred image.
A scheme of routinely selecting the best support is proposed. This scheme employs an objective error metric to evaluate the performance of Davey's blind deconvolution algorithm [Davey et al, Optics Comms., 69:356, 1989]. The effectiveness of this scheme for choosing the best support sizes as determined by testing is demonstrated.
Two new iterative algorithms are presented for solving blind deconvolution problems. Both algorithms incorporate a number of refinements into the Davey Algorithm. The performance of the first algorithm, referred to as the Automatic Iterative Algorithm (AIA), has been extensively examined in deconvolving images contaminated with different levels of Gaussian noise. The superiority of the AJA over the Davey Algorithm is demonstrated in terms of ease of parameter selection, computational efficiency and accuracy of reconstructions. The second algorithm referred to as the Coloured Noise Algorithm (CNA) is designed for deconvolving images contaminated with coloured noise. The superior performance of the algorithm over the Davey Algorithm in the presence of such noise is also demonstrated.
Various existing blind deconvolution techniques are reviewed. A comparison of the success of two techniques in solving the same problem is presented. The first new algorithm referred to above is chosen to represent iterative methods and is demonstrated to converge faster and produce better quality restorations for contaminated images than a representative optimization method (conjugate gradient).
Finally, computational simulations related to the contention that redundancy exists in multidimensional blind deconvolution are presented. The study suggests that blind deconvolution in three or more dimensions is overdetermined by sampling the convolution at the Nyquist rate.