Abstract:
In recent years, with increase in the use of Internet the multimedia contents on it have rapidly
increased. Users may want to go through a video in top down manner i.e. browsing the
videos, or in bottom up manner i.e. retrieving specific information from videos. They may
also want to go through the summary or through the highlights of the videos. Video data
is a major multimedia data available over the web; people want interactions to be possible
with the videos. This has necessitated the need to handle multimedia resources effectively.
Lecture videos are the category of videos that intrigue the users to interact with the videos.
This dissertation work proposes an automatic method for aligning scripts of lecture videos
with captions. Alignment is needed to extract time information from captions and insert it
in the scripts to create index of the videos. No alignment work has been previously done
in lecture videos domain. Alignment methods proposed for other type of videos are not
applicable for lecture videos because, different similarity techniques behave differently on
different types of datasets. The proposed method uses transcripts of lecture videos, SRT
file of captions available along with lecture videos and caption files generated from autocaption
generation feature of YouTube. The captions and scripts are then aligned using a
dynamic programming technique. No such work has been previously done for lecture videos.
Most important aspect of alignment is similarity measure. In the proposed work we have
used three similarity measures cosine, jaccard, and dice. A comparative analysis of these
measures is given in the dissertation. We also use a large lexical database of English words
known as WordNet for word-to-word similarity. The experimental result shows comparison
of accuracy of alignment for various similarity techniques and comparison of accuracy of
alignment for captions available along with lecture videos and captions generated from
YouTube’s auto caption generation feature.
