基于视觉语义概念的暴恐视频检测

摘要

暴恐视频在互联网的传播严重威胁着社会稳定与安全，对暴恐视频进行检测充满技术挑战。文章创建了用于评价暴恐视频检测算法性能的暴恐视频数据集，对数据集进行了语义概念和视频两个层面的标注。研究了基于灰度直方图、颜色直方图、颜色矩、局部二值模式及梯度方向直方图等特征描述子和支持向量机及极限学习机的视觉语义概念检测，针对不同的视觉语义概念通过仿真实验确定了最佳特征与分类器组合方式及相应参数设置。提出了基于灰度质心的视频关键帧提取方法，利用帧间相似性和视频序列相似性判别完成镜头突变和渐变判别，完成视频关键帧的提取。基于视觉语义概念和视觉语义概念词袋构建暴恐视频词频特征，使用支持向量机和视觉语义词频特征完成了暴恐视频内容的检测，并通过实验验证了该方法的可行性。%The spreading of terrorist video on internet threated the public safety a lot, and the detection of terrorist video content is full of technical challenges. In this paper, a terrorist video dataset was constructed for evaluation of algorithms, and the dataset is annotated by visual semantic concepts. Methods based on five kind of feature descriptor (gray histogram, color histogram, color moment, local binary pattern and histogram of orientation of gradient) and support vector machine and extreme learning machine were studied for visual semantic concept detection. A video key frame extraction algorithm based on gray massive center was implemented, and shots gradual change and sudden change were detected using similarity between neighboring frames and video sequence. A terrorist video detection framework was proposed combined visual semantic concepts and bag of visual semantic concepts, and the result of simulation experiment proved the effectiveness of it.