Frame clustering technique towards single video summarization

Abstract

Recent advances in technology, multimedia and social networking sites have led to a massive growth in web video content available for the general population. This results in information overload and management problem of the same. In this context, video summarization plays an important role as it aims to reduce the content size of video and yet present the important semantic concepts in the video. This gives an opportunity to reorganize video content in most succinct form for efficient and on- demand user consumption. Video summarization in its true sense is a hard problem as it involves domain specific semantic understanding of video content and user expectations. Most of the existing approaches relies on segmenting video into contiguous shots & selecting one or more keyframes from each shot and present these keyframes as summary. Such approaches may work well if independent concepts in video appear only once. However, in videos where same concepts are repeated multiple times, these existing approaches may pick repeating summary frames belonging to same concepts. In this paper, we present a novel frame clustering approach for generating very concise summaries by grouping all frames of similar concepts together irrespective of their occurrence sequence. The proposed approach is aimed towards large videos in domains like travel guide, documentaries, dramas where video revolves around few repeating concepts. The approach utilizes multiple video features in a generic way for frame-similarity determination and is extensible for multi-video summarization. Experimental comparative results substantiate the efficiency of the proposed approach in generating concise video summaries on videos with repeating concepts.