| Περίληψη: | Electric Network Frequency (ENF) fluctuations are used as an emerging
technology for multimedia authentication, timestamping and geolocation in
forensics, for video sunchronisation in the film making industry etc. The employed
techniques require extracting fluctuations from the recordings in question and
comparing them to ground-truth frequencies.
The objective of this thesis is the study of different methods and factors affecting
the ENF extraction from the visual part of videos. For this purpose, three algorithms
were implemented, each one applied to a different video category. The first two, for
static global and rolling shutter videos respectively and the last one for non-static
global shutter videos.
The algorithms for the static cases were presented on white wall videos for
simplicity reasons but they were eventually expanded to colorful scenes. Multiple
factors can affect the process of ENF extraction so the influence of video compression
and type of lightbulb were investigated for both types of shutter. For the non-static
videos, a superpixel-based approach was used and different numbers of superpixels
were tested in order to determine the least amount needed for the success of the
algorithm.
The validity of the algorithms was checked by correctly identifying the video’s
recording time for all the examined cases. According to the findings of the
experiments the results vary depending on the type of shutter. Global shutter videos
maintain the same similarity index between the ENF and the ground-truth signal for
all kinds of compression when for rolling shutter videos the index is significantly
more impacted. Additionally, when it comes to room illumination while incandescent
lights produce the finest results for rolling shutter videos, LEDs perform best for
global shutter videos. The superpixel technique works for both static and non-static
videos, and the more movement is incorporated into the video, the more superpixels
are required for the correct timestamping.
|
|---|
