There has been a growing interest in reversible watermarking of medical images re- cently for security reasons. Typically, humans are assumed to be the end user of watermarked images, however in many cases machine vision processes may be addi- tional consumers. Therefore, any watermarking performed on these images must be imperceptible to not only human users, but also these machine vision processes. The objective of this thesis is to understand the extent to which reversible water- marking affects the ability of computer vision algorithms to perform correctly. We address both the effect on primitive feature detection and on complete machine vi- sion processes, and investigate the ability to predict these effects using Image Quality Metrics (IQMs). Additionally, we describe the development of a new watermarking algorithm. We perform primitive feature detection on original and watermarked images, com- paring the output feature maps. Subsequently we use statistical modelling to allow prediction of feature map differences based on various IQMs of a watermarked image. We then conduct a similar experiment using manually specified feature maps and edge detectors across their full parameter space. Watermarking algorithms showing the least impact are highlighted and prediction of poorer performance a priori is investigated. In many cases watermarking is shown to cause a significant difference in the output feature map, however prediction of the difference is possible with excellent discrim- ination in many cases. A validation system for utilising these results in practical applications is presented. Three machine vision processes are investigated using a range of watermarking al- gorithms and embedding capacities – iris recognition, medical image registration, and diabetic retinopathy assessment. Significant differences are found in some cases, however at low capacities the iris and retinopathy processes show no significant dif- ferences. In addition, prediction of erroneous results for the retinopathy process was possible with excellent discrimination.
| Identifer | oai:union.ndltd.org:bl.uk/oai:ethos.bl.uk:619175 |
| Date | January 2013 |
| Creators | Lock, Andrew |
| Publisher | University of Aberdeen |
| Source Sets | Ethos UK |
| Detected Language | English |
| Type | Electronic Thesis or Dissertation |
| Source | http://digitool.abdn.ac.uk:80/webclient/DeliveryManager?pid=214152 |
Page generated in 0.0024 seconds