基于排序学习的视频摘要

Video summarization based on learning to rank

视频数据的急剧增加,给视频的浏览、存储、检索等应用带来一系列问题和挑战,视频摘要正是解决此类问题的一个有效途径。针对现有视频摘要算法基于约束和经验设置构造目标函数,并对帧集合进行打分带来的不确定和复杂度高等问题,提出一个基于排序学习的视频摘要生成方法。该方法把视频摘要的提取等价为视频帧对视频内容表示的相关度排序问题,利用训练集学习排序函数,使得排序靠前的是与视频相关度高的帧,用学到的排序函数对帧打分,根据分数高低选择关键帧作为视频摘要。另外,与现有方法相比,该方法是对帧而非帧集合打分,计算复杂度显著降低。通过在TVSum50数据集上测试,实验结果证实了该方法的有效性。

The exponential increase in the number of online videos has resulted in several challenges as regards video browsing,video storing,and video retrieval.These challenges can be effectively solved by video summarization.The existing video summarization methods construct objective functions based on empirical constraints and experience setup resulting from scoring for a set of frames.However,these methods have uncertainty and high complexity;therefore,in this paper,a video summarization method based on learning-to-rank algorithm is proposed.The proposed method considers summary extraction as a correlation ranking problem between frames and video.First,the training set is used to learn the ranking function,which places the frames having high correlation with video in the front position.Then,the score of each frame is calculated using the learned ranking function.Finally,the keyframes with high scores are selected as the video summary.Compared with the existing methods,the proposed method calculates a score for each frame rather than for a set of frames;therefore,computation complexity remarkably decreases.In addition,the effectiveness of the proposed approach is validated using experimental results on TVSum50 dataset.