Claxton, Christopher David (2007) Colour depth-from-defocus incorporating experimental point spread function measurements. PhD thesis, University of Warwick.

Preview

PDF WRAP_THESIS_Claxton_2007.pdf - Requires a PDF viewer such as GSview, Xpdf or Adobe Acrobat Reader Download (21Mb)

Abstract

Depth-From-Defocus (DFD) is a monocular computer vision technique for creating
depth maps from two images taken on the same optical axis with different intrinsic camera
parameters. A pre-processing stage for optimally converting colour images to monochrome
using a linear combination of the colour planes has been shown to improve the
accuracy of the depth map. It was found that the first component formed using Principal
Component Analysis (PCA) and a technique to maximise the signal-to-noise ratio (SNR)
performed better than using an equal weighting of the colour planes with an additive noise
model. When the noise is non-isotropic the Mean Square Error (MSE) of the depth map
by maximising the SNR was improved by 7.8 times compared to an equal weighting and
1.9 compared to PCA. The fractal dimension (FD) of a monochrome image gives a measure
of its roughness and an algorithm was devised to maximise its FD through colour
mixing. The formulation using a fractional Brownian motion (mm) model reduced the
SNR and thus produced depth maps that were less accurate than using PCA or an equal
weighting. An active DFD algorithm to reduce the image overlap problem has been
developed, called Localisation through Colour Mixing (LCM), that uses a projected colour
pattern. Simulation results showed that LCM produces a MSE 9.4 times lower than equal
weighting and 2.2 times lower than PCA.
The Point Spread Function (PSF) of a camera system models how a point source of
light is imaged. For depth maps to be accurately created using DFD a high-precision PSF
must be known. Improvements to a sub-sampled, knife-edge based technique are presented
that account for non-uniform illumination of the light box and this reduced the
MSE by 25%. The Generalised Gaussian is presented as a model of the PSF and shown to
be up to 16 times better than the conventional models of the Gaussian and pillbox.

Item Type:	Thesis or Dissertation (PhD)
Subjects:	T Technology > TA Engineering (General). Civil engineering (General)
Library of Congress Subject Headings (LCSH):	Computer vision
Official Date:	January 2007
Dates:	Date Event January 2007 Submitted
Institution:	University of Warwick
Theses Department:	School of Engineering
Thesis Type:	PhD
Publication Status:	Unpublished
Supervisor(s)/Advisor:	Staunton, Richard C., 1952-
Sponsors:	University of Warwick
Extent:	xxi, 265 leaves
Language:	eng
URI:	http://wrap.warwick.ac.uk/id/eprint/4461

Request changes or add full text files to a record

Actions (login required)

View Item

Downloads