- Library Home /
- Search Collections /
- Open Collections /
- Browse Collections /
- UBC Theses and Dissertations /
- A neural network implementation for integrating discontinuity...
Open Collections
UBC Theses and Dissertations
UBC Theses and Dissertations
A neural network implementation for integrating discontinuity and displacement information Ralph, Scott
Abstract
An artificial neural network is presented which performs the task of integrating discontinuity and displacement information to produce an estimate of displacement which is robust in portions of the image containing discontinuous motion. The behavior of the algorithm is learned from a series of randomly generated examples, and demonstrates that the necessary constraints required to perform fundamental visual computation can be extracted directly from a series of random images. By phrasing the problem as a machine-learning problem we are able to give explicitly the optimality criteria (given by the learning rule), without making explicit assumptions on the image features necessary to perform the computation. The motion primitives consist of the sum-of-squared-differences (SSD) values over a set of oriented rectangular support regions. A set of related measures, called shear values, are used to detect the position and orientation of discontinuities in the displacement field. By using a set of support regions which vary in shape and size, the algorithm is able to exploit the discontinuity information and choose the support region which best captures the underlying motion of the region. The resulting algorithm is compared to the traditional SSD algorithm with a single square support region, using both natural and synthetic images. Analysis of the algorithm indicates the neural network is able to reduce the distortion effects occurring near discontinuities and produces object boundaries which are significantly better representations of the object's true structure. The computed displacements show the neural network is able to interpolate over the SSD surface to produce displacements which are within sub-pixel accuracy. Additional confidence measures are given for the neural network and are compared to the traditional SSD algorithm.
Item Metadata
Title |
A neural network implementation for integrating discontinuity and displacement information
|
Creator | |
Publisher |
University of British Columbia
|
Date Issued |
1991
|
Description |
An artificial neural network is presented which performs the task of integrating discontinuity and displacement information to produce an estimate of displacement which is robust in portions of the image containing discontinuous motion. The behavior of the algorithm is learned from a series of randomly generated examples, and demonstrates that the necessary constraints required to perform fundamental visual computation can be extracted directly from a series of random images. By phrasing the problem as a machine-learning problem we are able to give explicitly the optimality criteria (given by the learning rule), without making explicit assumptions on the image features necessary to perform the computation.
The motion primitives consist of the sum-of-squared-differences (SSD) values over a set of oriented rectangular support regions. A set of related measures, called shear values, are used to detect the position and orientation of discontinuities in the displacement field. By using a set of support regions which vary in shape and size, the algorithm is able to exploit the discontinuity information and choose the support region which best captures the underlying motion of the region.
The resulting algorithm is compared to the traditional SSD algorithm with a single square support region, using both natural and synthetic images. Analysis of the algorithm indicates the neural network is able to reduce the distortion effects occurring near discontinuities and produces object boundaries which are significantly better representations of the object's true structure. The computed displacements show the neural network is able to interpolate over the SSD surface to produce displacements which are within sub-pixel accuracy.
Additional confidence measures are given for the neural network and are compared to the traditional SSD algorithm.
|
Genre | |
Type | |
Language |
eng
|
Date Available |
2010-12-02
|
Provider |
Vancouver : University of British Columbia Library
|
Rights |
For non-commercial purposes only, such as research, private study and education. Additional conditions apply, see Terms of Use https://open.library.ubc.ca/terms_of_use.
|
DOI |
10.14288/1.0051675
|
URI | |
Degree | |
Program | |
Affiliation | |
Degree Grantor |
University of British Columbia
|
Campus | |
Scholarly Level |
Graduate
|
Aggregated Source Repository |
DSpace
|
Item Media
Item Citations and Data
Rights
For non-commercial purposes only, such as research, private study and education. Additional conditions apply, see Terms of Use https://open.library.ubc.ca/terms_of_use.