Scene alignment by SIFT flow for video summarization

Video summarization is an efficient and flexible way to represent video data. In this paper, we use the kernel PCA and clustering based key frame extraction to realize multilevel video representation. In order to remove the redundancy caused by large scene changes, SIFT flow scene alignment is performed on the clustering set of key frames. After alignment, one representative frame is chosen from the reconstructed cluster set on matched frame pairs. We explore the difference on data structures between frame level and scene level, and modify the FCM method on the cluster number initialization for video summarization. Experimental results are presented to verify the efficiency of our approach.