Linear analysis of incidence structure: applications in non-rigid object recognition
Publication Date
October 15, 2019
Creator
Abstract
Experimental developments in object recognition systems are delivered through image analysis. In non-rigid object recognition, view-invariance is a limiting factor in algorithm development. For supervised learning and parametric training data sets, recognition often fails when the statistics of features are not accounted for or are not discriminative. It is therefore imperative to identify, and appropriate salient features, input to the object recognition system. Features undergoing rigid body and projective transformations exhibit nonlinear relationships and as such, interactions between features can be complex to identify and measure. This thesis contains research into incidence geometry applied to non-rigid object recognition. The first part of this work measures correlation accuracy to recover angular displacements. Using a combination of nonuniform sampling and up-sampling, a matched filter analysis reveals a relaxation in interpolation complexity between sampled grid points. Furthermore, the technique identifies application in natural image structure analysis. The second part investigates the accuracy and precision of sub-pixel edge feature measurements. An arbitrary edge direction detection method based on non-integer coefficients and quadratic refinement reveals a precision measurement perhaps applicable for medical and manufacturing image screening applications at the millimetre and micron scale. Lastly, a recognition process utilising the Hough transform to measure and accumulate critical object feature statistics is investigated. Based on the training examples used to test the approach, the recognition error of a front profile was 8.2%. Identified sources of error include the profile measurement point locations and their approximation to characterise a profile. The results and applications of the novel adapted signal processing techniques are examined.
Item Type
ethesis
Thesis Type
PhD
Subjects (LC)
Associated Schools / Departments
Faculty of Engineering
eprints ID
57192
UoN Repository URI
Except where otherwise noted, this item's license is described as
File(s)![Thumbnail Image]()
Name
4224685_N_Wells_PhD_Thesis.pdf
Type
Full-text
Description
Examined
Size
7.76 MB
Format
Adobe PDF
Checksum (MD5)
d1c35e8b665097b1dc751284214ad385