A Noise-Reduction Approach to Scene Segmentation for Large Video Databases

Automatic video segmentation is the first and necessary step that structures a video into several smaller and meaningful units for effective browsing and retrieval for large video databases. The effectiveness of this step is, thus, very crucial to the overall performance of a video database management system. In this paper, we present a novel concept in scene segmentation called Noise-Reduction Scene Segmentation. This approach discards irrelevant areas or noise in a video frame from being used in the segmentation process to increase the accuracy of the segmentation. Unlike existing techniques, video frames are first noise-reduced and only relevant information is left for subsequent steps of the segmentation process. Our experimental results indicate that a seamless integration of our simple noise filter to an existing scene segmentation technique offers a non-negligible improvement in the segmentation accuracy (i.e., as much as 59% less falsely detected scenes).