基于机器学习的压缩域镜头分割技术

Shot Segmentation Technology by Using Machine Learning on Compressed Domain

为了进行视频结构化和视频内容分析,需要准确有效地提取视频镜头的边界信息。为此提出了一种利用支持向量机(SVM)学习压缩域特征的算法进行镜头边界检测,只需简单译码即可得到MPEG1/2等各类视频流压缩域的特征信息。经TRECVID2005镜头边界检测集的评测,该算法在保证查全率和检测精度的情况下获得了满意的效果。

To analyze the video structure and content, shot boundary features should be extracted exactly and efficiently. In this paper, a novel approach is proposed to detect shot boundaries by using Support Vector Machine（SVM） on the compressed domain features, which can be extracted without fully decompressing MPEG - 1/2 video. Experimental results on TRECVID2005 shot boundary evaluation demonstrate that the proposed approach can obtain promising performance under good recall and precision.