面向暴恐音视频的内容检测系统研究与实现

Research and Implementation of Content Detection System for Violent Video

网络上有海量的音视频,其中暴恐音视频不容忽视。对面向暴恐音视频的内容检测系统进行研究,选取音频特征MFCC作为检测特征,采用词袋模型建模,利用支持向量机分类,优化了多个全局参数,过滤了无用镜头,采用欠采样平衡正负样本的数目差距,减少了冗余和训练复杂度,取得了理想的检测效果,且能给出暴恐程度的评估。具体地,提出了词频加权参数c的改进算法和基于距离倍数的词频加权改进算法,能使词袋模型表示更精确,从而提高了准确率。

Lots of audio and video exist on the Internet, of which the violent audio and video cannot be ignored. The content detection system of violent audio and video is discussed. The audio feature MFCC is selected as the detection feature, the word-bag model used for system modelling and the support vector machine （SVM） in classifying and optimizing the multiple global parameters. Meanwhile the useless len is filtered out, and the number gap of between the positive sample and negative sample balanced by under sampling, thus to reduce the redundancy and training complexity, achieve the desired detection effect, and give the degree evaluation of the violent terror. Specifically, the modified algorithm of word-frequency weighted parameter C and the modified word-frequency weighting algorithm based on distance multiplier are proposed, thus to make the word-bag model more precise and improve the accuracy.