Content-Adaptive Video Summarization Combining Queueing and Clustering

This paper presents an efficient three-step algorithm to compute key frames for summarization and indexing of digital videos. In the preprocessing step, we remove the video frames that constitute the gradual transitions of video shots using an entropy-based energy-minimization method. The second step measures the content dissimilarity of video frames and removes the ones that are similar in content. In the postprocessing step, the video frames are processed using a dynamic clustering technique. The first and the second steps are implemented as queues, allowing efficient temporal filtering of video frames. Our algorithm greatly reduces the difficulty of parameter selection with content-adaptive parameters. Experimental results on four videos show that our method retrieves the least number of transitional and near-duplicate key frames, compared with three other existing methods.