Methods for enhancing underwater imagery
Publisher
Metadata
Show full item recordAbstract
This thesis explores underwater imaging solutions in order to alleviate the poor contrast and
the distortion in the perception of colour caused by the processes of scattering and absorption.
We demonstrate through simulated experiments that imaging systems with higher spectral
resolution than RGB could be useful for underwater imaging tasks such as estimating illumination
and spectral reflectance values. We also tested hyperspectral imagers in real world experiments
and found that the current technology is limited for underwater image enhancement applications.
To address the problem of poor visibility in underwater scenes we introduce dehazing methods
for underwater RGB images and videos. Current underwater dehazing methods suffer from
limitations such as estimated parameters being biased towards pixels of bright objects in a scene
and artefacts being created in regions that contain pure haze. Bright objects in a scene are avoided
by using texture features during the estimation of parameters and local bias is avoided by taking
information from an image at different spatial resolutions. We inhibit noise and artefacts being
created in the pure haze regions by segmenting these areas and treating them as a special case. We
address the spectral distortion present in underwater scenes by applying a water-type dependent
white balancing step. We also demonstrate the application of our method to underwater videos
with a weighted temporal smoothing of the estimated parameters and a Gaussian normalisation
step that ensures segmentation of pure haze regions is stable across frames.
We evaluate our methods both on quantitative metrics and through subjective experiments and
demonstrate an improved performance in comparison to the state of the art in underwater image
and video enhancement. We also show how no-reference underwater image quality assessment
metrics do not always correspond with human judgement and provide suggestions on how they
could be improved.
Authors
Emberton, SimonCollections
- Theses [4203]