Real-time keyframe extraction towards video content identification

Keyframe extraction constitutes a fundamental unit in many video retrieval-related applications. In the emerging research field of content-based video copy detection, efficient representation of video content at keyframe level is crucial, due to the fact that similarity search is mainly performed between content-representative frames. In this paper a sequential search algorithm that bypasses the process of temporal video segmentation is proposed for keyframe extraction in MPEG videos. We aim at providing an efficient, real-time and fully-automatic way of extracting keyframes in videos, where not only the laborious task of offline video database indexing is avoided, but also query video processing is performed in the same manner as the reference video database. Significant reduction in computational cost is achieved by exploiting DCT coefficients in feature extraction. The effectiveness of the proposed scheme is evaluated in terms of quality and speed on the manually annotated TRECVid 2007 test video dataset.