Visibility Recovery on Images Acquired in Attenuating Media. Application to Underwater, Fog, and Mammographic Imaging
Loading...
Identifiers
Publication date
2015-12-17
Authors
Advisors
Journal Title
Journal ISSN
Volume Title
Publisher
UPV-EHU
Citations
Export
Abstract
When acquired in attenuating media, digital images often suffer from a
particularly complex degradation that reduces their visual quality, hindering
their suitability for further computational applications, or simply
decreasing the visual pleasantness for the user. In these cases, mathematical
image processing reveals itself as an ideal tool to recover some
of the information lost during the degradation process. In this dissertation,
we deal with three of such practical scenarios in which this problematic
is specially relevant, namely, underwater image enhancement, fog
removal and mammographic image processing. In the case of digital mammograms,
X-ray beams traverse human tissue, and electronic detectors
capture them as they reach the other side. However, the superposition
on a bidimensional image of three-dimensional structures produces lowcontrasted
images in which structures of interest suffer from a diminished
visibility, obstructing diagnosis tasks. Regarding fog removal, the loss
of contrast is produced by the atmospheric conditions, and white colour
takes over the scene uniformly as distance increases, also reducing visibility.
For underwater images, there is an added difficulty, since colour is not
lost uniformly; instead, red colours decay the fastest, and green and blue
colours typically dominate the acquired images. To address all these challenges,
in this dissertation we develop new methodologies that rely on: a)
physical models of the observed degradation, and b) the calculus of variations.
Equipped with this powerful machinery, we design novel theoretical
and computational tools, including image-dependent functional energies
that capture the particularities of each degradation model. These energies
are composed of different integral terms that are simultaneously
minimized by means of efficient numerical schemes, producing a clean,
visually-pleasant and useful output image, with better contrast and increased
visibility. In every considered application, we provide comprehensive
qualitative (visual) and quantitative experimental results to validate
our methods, confirming that the developed techniques outperform other
existing approaches in the literature.