AbstractMany animals perceive changes in the polarization of light in addition to, or instead of, changes in intensity and wavelength, allowing them to more effectively perform tasks. Although humans lack this ability, recent work allows humans to exploit polarization using imaging polarimeters, and many applications which use this technology have been developed. A major challenge of polarimetry is image degradation due to noise, which has led to incorrect conclusions in the literature and reduces the effectiveness of computer vision algorithms.
Local feature extraction is a technique for extracting information from images which is a common intermediate step in many computer vision applications. No work has yet been done assessing the performance of existing local feature extraction algorithms with polarimetry, or how they can be most effectively used.
In this thesis degradation caused by noise in polarimetry is investigated, and mitigating steps are proposed. Denoising algorithms are then investigated which are shown to improve the peak signal-to-noise ratio of polarization images by 4.5dB over existing algorithms. This is done by adapting an existing denoising algorithm, Block-Matching 3D, to create a method specifically
for polarimetry, Polarization-BM3D (PBM3D). PBM3D will be shown to provide superior visual quality to existing algorithms.
This thesis also investigates the use of common local feature extraction algorithms with polarimetry, and compares their effectiveness with colour imagery. It will be demonstrated that using local features with polarimetry can yield better results than using colour imagery (by 7.5%), specifically when the Stokes representation is used. It will also be shown that Hessian-affine is
the most effective detection algorithm and SIFT is the most effective description algorithm for use with polarimetry. Finally the effects of noise on features extraction with polarimetry will be presented, and it will be shown that PBM3D improves detector and descriptor performance by 35% and 5% respectively.
|Date of Award
|25 Sept 2018
|Nicholas W Roberts (Supervisor) & David R Bull (Supervisor)