From local to global key-frame extraction based on important scenes using SVD of centrist features

Abstract

The wide spread of multimedia applications and the rapid growth of digital video data require efficient video summarization methods. Extracting brief and pertinent information allows users to quickly browse, recognize and understand a large amount of video content. In this paper, a video summarization method based on important scenes is proposed. A local selection of potential candidate key-frames (PCK) is first performed using only one iteration of the k-means algorithm, where its initialization is achieved using a dictionary selection. Scores of importance are calculated for each PCK to accomplish the global selection. While some approaches remove redundant key-frames to share unique information, this can be used to classify scenes by duration and temporal position. Following such classification, the scene with the longest duration can be considered as the most important one. Therefore, rules of insertion are defined to allow redundancy when the information is considered important. In our contribution, to represent a frame, the singular value decomposition (SVD) of centrist are used as features. The SVD of Centrist allows to better measure the similarity between adjacent frames than other features, and thus to enhance the performance. Experimental results over two different databases show the diversity of our summary and the effectiveness of our method compared to related state of the art methods.