Adaptive Canny Edge Detection: Hysteresis Thresholds with Deep Learning
Keywords
Loading...
Authors
Issue Date
2021-02-04
Language
en
Document type
Journal Title
Journal ISSN
Volume Title
Publisher
Title
ISSN
Volume
Issue
Startpage
Endpage
DOI
Abstract
This thesis proposes a novel deep learning-based approach
for adaptive Canny edge detection (CED).
Edge detection is an essential component of many computer
vision tasks. Prosthetic vision for example, aims
to restore vision in patients with the help of an implant
which converts digital information into brain signals. To
improve this technique we aim to automate the threshold
selection usually done manually for such an edge
detection application. This is accomplished by exploring
the possibility of a deep neural network learning the
optimal hysteresis threshold values to use in Canny edge
detection for a given input image.
The approach in this thesis consists of a prediction model
which is trained using a three-model pipeline with a
surrogate network and a validation model to select the
optimal threshold values for a given image.
The surrogate model shows sensitivity to different
threshold-sets similar to the CED and thus provides a
differentiable substitute. A classifier and an autoencoder
show good initial results in giving a performance indication
on the resulting contour-image. These results
open up the possibility to train a range of prediction and
validation model combinations on the surrogate model,
and give an indication of which validation methods
should be explored in the future.
Description
Citation
Faculty
Faculteit der Sociale Wetenschappen