基于标签分布学习的视频摘要算法

针对现有监督视频摘要算法中存在的模型训练复杂问题,提出一种新的基于标签分布学习(LDL)的视频摘要算法,采用非参数监督学习的方式生成视频摘要,利用标签传递的方法将摘要结构从带有注释的视频转移到相同类型的测试视频中.首先提取视频的卷积神经网络特征和颜色特征,将两者融合后进行降维得到特征矩阵;然后将特征矩阵与训练样本的标签分布一起输入到LDL模型中;最后根据模型输出的标签分布选取关键帧,生成视频摘要.在基准数据集上与其他算法的实验表明,该算法生成的摘要与用户创建的摘要一致性很高,明显优于其他算法.

Label Distribution Learning for Video Summarization

There is a problem of complicated model training in the supervised video digest algorithm.To solve this problem,a new video summary algorithm based on label distribution learning (LDL) is proposed.This algorithm uses non-parametric supervised learning to generate summarization.The main idea is to transfer summary structures from the annotated video to the same type of test video by label passing.Firstly,the convolutional neural network features and color features of the video are extracted.A feature matrix is obtained by combining these two features and reducing the dimension.It is then entered into the LDL model along with the label distribution of the training samples.Finally,the key frames are selected according to the label distribution of the model output,and they are composed into a video summary.By comparing the experiments with other algorithms on the benchmarks,it shows that the summaries generated by this algorithm are highly consistent with the human- created abstract,which is obviously superior to other methods.