Automatic Commercial Detection Based on Joint Video and Audio Features

• Main Authors: Cyong-Wun Fan, 范瓊文
• Other Authors: Shih-Hsuan Yang
• Format: Others
• Language: zh-TW
• Published: 2011
• Online Access: http://ndltd.ncl.edu.tw/handle/mg9fsp

碩士 === 國立臺北科技大學 === 資訊工程系研究所 === 99 === Broadcast TV programs intervene with commercials. Although commercials are important for advertiser and TV broadcasters, they will reduce the smoothness for watching TV programs. Detection of TV commercials is thus important for many applications. In this thesis, we propose a new TV commercial detection algorithm for drama programs in Taiwan. The proposed method integrates audio and video features to identify the starting and ending instants of program segments, and the starting and ending instants of commercial intervals. The employed visual characteristics include still pictures, rating logos, and program titles. The employed audio characteristics include the starting music, ending music, and short silence. We use program title comparison, frame difference, SVM (support vector machine), LSTER (low short-time energy ratio), and amplitudes of audio samples to detect the program title disappearing, still pictures, rating logos, starting music, ending music, and short silence. Then, we incorporate the information of still pictures, ending music, and short silence to decide the time instants of ending program segments and starting commercials. Finally, we use rating logos, starting music, and short silence segments to determine the time instants of ending commercials and starting program segments. Experiments show that the proposed method can detect all commercial intervals and the detected time instants are mostly within one second to the correct time instants.