Approche d’apprentissage pour l’analyse des Big Data
No Thumbnail Available
Date
2023
Authors
Benoughidene Abdelhalim
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
The development of information technology has led to the big data revolution, with the amount
of data produced increasing at a high rate. Video data is a significant component of big data,
and the concept of automatically analyzing this rapidly increasing video content has become a
popular research topic. In such a scenario, video automatic analysis uses a new generation of
information technologies, such as artificial intelligence (machine learning), which help transform
traditional video analysis to be more efficient and convenient.
Video summary (VS) is now one of the primary areas of study in video analysis. Despite
the use of big data-driven models, producing accurate video summaries in an efficient and effective
manner remains a challenging task. The most effective and efficient way to transform
lengthy, unstructured videos into structured, condensed, understandable, and useful information
is through the use of video summaries. The primary goal of summarizing a video is to
break it down into shots and select key frames from each shot that best capture the essence
and flow of the entire video.
The goal of this thesis is to improve the performance of video summarization systems by
enhancing the quality of analytical techniques. To achieve this goal, we proposed two contributions.
First, we suggested a shot boundary detection (SBD) method to adapt key shots
and exploit its potential in video summaries. This is the first step in the video summarization
process, and the results have a significant impact on the quality of the final summary. The
main idea behind SBD is to extract features from video frames and then identify the boundaries
between shots based on the differences in the features. Second, we focused on improving video
summaries using unsupervised machine learning techniques (DBSCAN) and genetic algorithms
(GA) to optimize DBSCAN hyperparameters. We validated the proposed methods and results
obtained through extensive comparative analysis using datasets.