Abstract: Content-based video analysis calls for efficient and flexible video representation. In this paper, a novel multi-resolution video representation (MRVR) scheme is proposed and realized by performing the kernel principal component analysis (KPCA) on the low-level visual features extracted from a video sequence. By simply changing the kernel parameters or the dimensionality of the subspace, this scheme can represent video content from coarser to finer levels in the subspace, according to its intrinsic structure. An application of keyframe extraction is investigated to show the advantages of this representation scheme. Furthermore, based on this scheme, a two-level video summarization approach is proposed to represent long video sequences. The experimental results of both short and long video sequences have demonstrated the effectiveness and flexibility of the proposed video representation scheme.