视频摘要的算法研究

发布时间：2018-04-23 17:26

本文选题：视频摘要 + 自适应　；参考：《杭州电子科技大学》2017年硕士论文

【摘要】：近年来,飞速发展的互联网和多媒体技术导致多媒体信息数据骤增。数字视频,作为主要的多媒体信息载体,广泛地应用于生活的方方面面。大量的视频一方面可以使人们以更丰富的形式获取信息,给生活带来了很大的便利。而另一方面,却给视频存储、传输、归档和检索带来巨大压力。因此,视频摘要技术应运而生。类似于文本摘要,视频摘要是对原始视频内容的总结和概括,通过分析原始视频数据流,从中选择有意义的视频内容组成紧凑的摘要。视频摘要可以结合视频标注技术用于视频检索;还可以形成独立的产品比如电影预告片,应用在我们的日常生活中。技术是目前计算机视觉领域的研究热点,传统的基于数据聚类的生成视频摘要的算法存在以下两点不足:第一,无法根据输入的不同时长和不同类型的视频自适应的得到最优的聚类数目。第二,在算法中仅提取了图像的颜色特征,忽略了图像的纹理和形状特征,而单特征既不能全面的表达图像的视觉信息,又不能有效的消除图像的噪声,导致生成的视频摘要的质量低下。针对以上不足,本文提出了一种基于自适应最优聚类与多特征融合的视频摘要生成算法。把视频分解为图像序列后,首先做预采样处理。然后提取预采样帧的颜色、HOG和LBP特征,并把这三种特征融合在一起来表征一幅图像。接着使用零均值归一化互相关指标作为帧与帧之间相似性度量的标准,把连续相似的帧分割为若干个镜头,从而得到最优的聚类数目。接着使用改进的k-means++算法对所有的帧进行聚类,选取距离聚类中心最近的帧作为关键帧。最后,分别计算所有关键帧的颜色直方图和梯度方向直方图的归一化方差,过滤掉无意义的关键帧。本文以匹配率和错误率两个相互补充的指标来衡量视频摘要质量的好坏。TRECVID2007数据集上的实验结果表明,本文算法具有良好的鲁棒性,进一步提高了所生成的视频摘要的质量。
[Abstract]:In recent years, the rapid development of the Internet and multimedia technology led to a surge in multimedia information data. Digital video, as the main multimedia information carrier, is widely used in all aspects of life. On the one hand, a large number of videos can enable people to obtain information in a richer form, which brings great convenience to life. On the other hand, it puts great pressure on video storage, transmission, archiving and retrieval. Therefore, video summary technology emerged as the times require. Similar to text summary, video summary is a summary and summary of original video content. By analyzing the original video data flow, a compact summary is made up of meaningful video content. Video abstracts can be used in video retrieval in combination with video tagging technology, and can also form independent products such as movie trailers, which can be used in our daily life. The traditional algorithm of generating video summary based on data clustering has the following two shortcomings: first, The optimal number of clusters can not be obtained adaptively according to different input time and different types of video. Secondly, only the color features of the image are extracted in the algorithm, and the texture and shape features of the image are ignored, while the single feature can not only fully express the visual information of the image, but also can not effectively eliminate the noise of the image. The resulting video digest is of low quality. In this paper, a video summary generation algorithm based on adaptive optimal clustering and multi-feature fusion is proposed. After the video is decomposed into image sequences, presampling is done first. Then the color hog and LBP features of the presampled frames are extracted, and the three features are fused together to represent an image. Then the zero mean normalized cross-correlation index is used as the criterion of similarity measurement between frames. The continuous similar frames are divided into several shots and the optimal clustering number is obtained. Then the improved k-means algorithm is used to cluster all frames and the nearest frame is selected as the key frame. Finally, the normalized variance of color histogram and gradient histogram of all key frames are calculated, and the meaningless key frames are filtered out.
【学位授予单位】：杭州电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】