Abstract:
"
Rain streaks are considered as a disturbance or a noise which cause performance or
accuracy limitations in many computer vision systems. This obstructing rain streak
problem is addressed within the domain of “De-raining”. De-raining is the process of
removing rain streaks from media such as photos or videos. Two main sub-categories
of de-raining are video de-raining and single image de-raining. This work is based on
single image de-raining. Many other single image de-raining works have taken place
using both traditional and deep learning approaches. A critical analysis of novel,
credible, and best performing single image de-raining systems was done to identify a
research gap in the domain. Identified research gap is an unpaired training gap which
causes a performance limitation within current image de-raining systems due to the
unavailability of natural paired rain image data. The project DERAINIZER is a novel
single image de-raining framework which addresses this research gap by applying
image-to-image translation techniques. It enables future researchers of the domain to
employ unpaired training data to train their models. This work is evaluated by
benchmarking achieved results for popular datasets, against existing systems in the
domain.
"
