Unsupervised Classification of Videos Using Semantic Analysis (PhD Thesis)
- Karachi : NED University of Engineering and Technology Department of Computer Science and Information Technology, 2019
- XX, 134 p. : ill
Includes Bibliographical References
ABSTRACT :
Advertisements displayed in TV broadcast are a very important part of transmission as majority of revenue for a broadcaster is generated by advertising. Fast and accurate advertisement discovery is an important issue in research community of computer vision. The main challenge for advertisement detection is lack of legislation for media industry in Pakistan that ensures the separate identification of advertisements and other non-advertisement programs through blank frame or absence of TV channel logo during TV transmission. In this thesis a framework for TV Commercial Detection and Identification (UTCDI) is proposed. The framework is composed of two sections; 1.) Unsupervised detection of TV commercials. 2.) Recognition or identification of commercials. An algorithm has been developed for TV commercials (ad) detection using semantic analysis. The term semantic in video processing refers to the any high-level concept that human can understand such as soccer ball in sports video, lion in animal planet documentary program, a ship and a rock. In case of this research study, an ad is semantic feature, which is composed of other semantic notions such as shots and scenes. The proposed algorithm uses colour histogram feature to detect scene change points in video. Furthermore, these scenes are used for unsupervised advertisement detection. Algorithm assigns unique ID automatically to each scene; boundaries of each repeating pattern of scenes are identified. If the repeating patterns are repeated more than a threshold value then they are marked as a commercial and assigned an ad ID. This approach is termed as semantic advertisement discovery. As this approach solely relies on repeated scenes which is high-level concept rather than computing text, edge and shape of any object from each frame in result the approach saves the computation and time. This approach successfully results the unsupervised discovery of TV advertisement with average value of precision 97% and average recall rate is 85%. Meanwhile, it suffers a problem that the results heavily rely on the size of video file. The more the size of video data the higher would be the chance of repeated scenes, the shorter the size of video data the lesser would be the chance of advertisement discovery as it depends on the threshold value for repeated scenes. Another factor that significantly influences the advertisement detection is quality of video. The proposed algorithm does not provide good ad detection results on grainy and poor quality videos.
In the second section, The SURF (Speeded Up Robust Features) feature is computed of key object from candidate advertisement. Then the SURF feature and the colour histogram feature are used to obtain a train file - which in result recognize and identify target commercial in a TV transmission file.
The framework for ad discovery based on semantic segmentation of broadcasted TV transmission can be used for PEMRA (Pakistan Electronic Media Regulatory Authority) to identify a particular ad and its statistics. The framework is also usable for Advertisements agency to visualize or measure the air time used for different ads. Experimental results confirm the importance of the proposed framework.
Human Computer Interaction Thesis Semantic Computing Thesis TV ads Thesis Ad Detection Thesis