A SOM-based Approach to Video Summarization System

• Main Authors: Chien-Hang Chen, 陳劍航
• Other Authors: Mu-Chun Su
• Format: Others
• Language: zh-TW
• Published: 2006
• Online Access: http://ndltd.ncl.edu.tw/handle/3u2t3g

碩士 === 國立中央大學 === 資訊工程研究所 === 94 === Due to rapid advances and improvements in electronics hardware and networking technologies and the decreasing cost of storage, video data are becoming available at an ever increasing rate. Traditional database management technique for text documents cannot effectively data with video data; therefore, Method and technique to automatically analyze video data have become a very attractive and challenging research topic. An efficient video database management system should have following two functionalities: 1) the video summarization functionality which make the take of browsing video content become easy and 2) the video retrieval functionality which can retrieve video from a huge video database based on user queries. This thesis focuses on the development of video summarization technique. Traditional way to browse video data is via the “fast forward” and “rewind” function keys to manually locate the region of interest. It is very time consuming. The goal of the new video summarization technique proposed in this thesis is to provide and effective table of content which can capture the semantic structure of a vide document. Several different approaches to video summarization technique have been proposed, each for its own advantage and limitation. The proposed video summarization technique involves the following four steps: 1) shot detection, 2) key-frame extraction, 3) shot group, 4) scene detection. The most appealing property of the proposed technique is the use of the self-organizing feature map(SOM).Since the SOM has the topologically preserving property, shots with similar feature will be grouped into the scene cluster and similar will be located nearby on a map. Then a region growing technique is employed to merge similar shots into groups. After the group map has been constructed, an effective scene detection technique is adopted to merge groups with a similar semantic concept into a scene. The constructed scene map can be either directly used as the table of content of a video document or transformed to a hierarchy tree to represent the video content of a video document. Via the scene map or the hierarchy tree, a user can effectively browse the content. The performance of the proposed technique is demonstrated by experiments on several different types of video documents.