Audio-Assisted Scene Segmentation for Story Browsing

Content-based video retrieval requires an effective scene segmentation technique to divide a long video file into meaningful high-level aggregates of shots called scenes. Each scene is part of a story. Browsing these scenes unfolds the entire story of a film. In this paper, we first investigate recent scene segmentation techniques that belong to the visual-audio alignment approach. This approach segments a video stream into visual scenes and an audio stream into audio scenes separately and later aligns these boundaries to create the final scene boundaries. In contrast, we propose a novel audio-assisted scene segmentation technique that utilizes audio information to remove false boundaries generated from segmentation by visual information alone. The crux of our technique is the new dissimilarity measure based on analysis of statistical properties of audio features and a concept in information theory. The experimental results on two full-length films with a wide range of camera motion and a complex composition of shots demonstrate the effectiveness of our technique compared with that of the visual-audio alignment techniques.