【作者】 陈斌；

【导师】 蒋兴浩； 孙锬锋；

【作者基本信息】 上海交通大学 ， 通信与信息系统， 2010， 硕士

【摘要】 视频内容的自动分类算法是计算机视觉领域中一个重要的研究课题,它为日益增加的视频数据的管理提供了方便,基于内容的视频自动分类作为视频传播控制的一类关键技术在对网络媒体进行有序管理的需求下至关重要。基于视频自动分类技术的应用,媒体网站可以把海量的视频内容进行自动分类,从而进行更有效的组织、存储和检索,还可实现对不良视频信息,如恐怖暴力视频的自动初步筛选。视频自动分类算法的性能极大地依赖于视频特征的提取和分类算法的选取,本文从对视频内容和视频风格类型的角度出发,提出了基于视觉和音频双模态特征组合的视频特征提取方法,和改进支持向量机(SVM)视频分类算法,实现了对卡通、广告、音乐、新闻、体育这五类常见视频的自动分类,以及对电影中恐怖暴力场景的自动识别。首先,在分析现有的视频分类算法的基础上,针对现有算法存在的问题,通过分析五类常见视频在视觉上的差异,本文提出了新的特征表达方案即MPEG-7视觉描述子组合模型,从颜色、纹理、形状、运动四个方面提取了共九种描述子来构成新的整体视觉特征,取得了较好的效果;在识别恐怖暴力场景时,本文根据这些场景的特点采用了视觉和音频两种模态的特征,相比单一特征增加了场景模式匹配的准确率,在有效性和区分度上达到了满意的效果。在选择并提取了合适的特征后,针对目前统计方法中存在的通过小样本集很难设计有效分类器的问题,本文提出了基于支持向量机的视频自动分类算法,并对分类器的判决策略方法进行了改进,提出了基于支持向量机1-1方法的二次预测机制,进一步提高了支持向量机多分类方法的准确率。仿真实验的结果表明:本文算法在特征选择方面突出了不同类别视频的差异性,增强了待分视频的区分能力;其次,改进的二次预测机制提高了支持向量机的多视频分类的性能;最后,与目前的相关类似算法进行了对比实验,五类视频的分类实验和恐怖暴力场景的识别实验均证明了本文算法在视频分类准确率方面的优越性能。更多还原

【Abstract】 Automatic video classification is an important subject in the computer vision research area. It provides more convenience to the management of increasing video data nowadays, and is essentially crucial to the video broadcasting control technology. Based on the application of automatic video classification, the multimedia websites could classify the huge amount of video data automatically and easily, and further organize, store and index them more effectively. They could also use this technology to filter the bad video content primarily.The performance of video content classification largely relies on the video features and classifiers. This paper explores the problem from the view of video content and genres, presents the feature extraction method based on audio-visual features and the multi-class classification algorithms based on modified SVM.Based on the analysis of recent video classification algorithms, aiming at figuring out their problems, this paper presents a new feature combination model which includes both visual and audio features. Visual features consist of nine MPEG-7 visual descriptors, which include color, texture, shape and motion descriptors, while audio features consist of nine most useful time domain and frequency domain features. The features chosen enlarged the coherence of one class and difference of different classes, and reached a satisfied effect.After extracting the appropriate features, this paper describes the automatic video classification algorithm based on SVM and modifies the classify decision policies, namely, introduces the second-prediction strategy into the SVM 1-1 method, and improves the classification accuracy.Experiments show that: the features extracted have improved the identifiability of different video genres, and made them more easy to be classified; performance of the SVM classifier has been enhanced by introducing the second-prediction strategy; results compared with other video classification algorithms demonstrate the accuracy and effectiveness of this approach.更多还原