APP下载

基于多尺度时长音频特征的暴力音频事件检测

2014-04-29徐战苍李海峰

智能计算机与应用 2014年5期

徐战苍 李海峰

摘要:暴力镜头检测是近年来的研究热点之一。早期的暴力镜头检测主要依赖视频特征,由于音频信息具有良好的稳定性和在不同文化和人群之间的一致性,现在人们越来越多地关注音频信息的使用。本文研究使用音频特征对电影镜头中的暴力音频事件进行检测。本文提出了一种基于多尺度时长的特征提取方法。提取了除MFCC、LPC、能量等短时特征以外,还提取了能量均值方差、子带能量均值和方差、帧间差分等长时特征。暴力镜头中出现较多且具有代表性的音频事件有爆炸、尖叫、枪击3种。本文以电影的镜头为识别单位,使用支持向量机分类算法实现了一个检测系统。通过在15部好莱坞电影上的实验,表明本文基于多尺度时长的音频特征在暴力音频事件检测工作中,能够取得较好的结果。

关键词:暴力镜头检测;多尺度时长特征;音频事件检测;支持向量机

中图分类号: TP391 文献标识码: A 文章编号:2095-2163(2014)04-

Violent audio event detection based on multi-scale audio features

XU Zhancang,LI Haifeng

( School of Computer Science and Technology,Harbin Institute of Technology,Harbin 150001,China)

Abstract: Violence detection is one of the hot research topic in recent years. Early work mainly depends on the video characteristic, considering the audio information has good stability and consistency between different cultures and people, people are now more and more attention to the use of audio information. This paper studies using audio features to detect violent audio event in the movie. So this paper presents a multi-scale feature extraction method. The extracted features besides MFCC, LPC, short term energy also extracted the long term feature, such as the mean and variance of energy and sub-band energy, difference between frames. The audio events appeared frequently in violence scenes are explosions, screams, gunshots. Therefore, using support vector machine classification algorithm, the paper implements a detection system, to detect the violent audio event in the movie scenes. Through experiments on 15 Hollywood movies, experiments results show that the multi-scale audio features can achieve good results in the violent audio event detection work.

Key words: Violence Detection; Multi-scale Audio Features; Audio Event Detection; Support Vector Machine (SVM)

0引言

音视频中暴力与色情内容的自动分析与检测已然成为近年来研究工作的新鲜热点之一,而倍受各方关注。尤其在网络发达的今天,包含过多儿童不宜观看内容的音视频,如果只是依靠人工审核,即使耗费大量的人力及财力也难以即时发现不良信息。本文即针对电影中所包含的音频暴力片段进行研究,提出了基于音频长短时特征与统计特征相融合的组合量特征,由此而实现了一种基于音频的长短时特征与统计特征融合的暴力内容自动分析系统。……

登录APP查看全文