网络敏感视频识别的关键问题研究

The emergence and development of video-sharing sites, P2P, microblog, podcasting, mobile WAP sites and 3GP sites facilitate the dissemination of sensitive videos, which poses a major threat to national security, social stability, and physical and mental health of teenagers. With the increasing demands for network content security, this project is intended to explore some important frontier problems of web sensitive video recognition. For recognizing specific types of sensitive videos, this project emphasizes on two aspects: pornographic video recognition based on probability fusion from multi-modals and multi-clues; violent video recognition via incorporating contextual information from images, motion analysis, audio intensity, emotional expression, etc. For customizable instance- driven sensitive video recognition, this project focuses on sensitive video recognition based on specific target attribute learning and that based on multi-feature semantic association copy detection. Eventually, a web sensitive video recognition system shall be completed. The project will result in several invention patents and other independent intellectual property rights. Meanwhile, a series of high-quality papers shall be published in international top journals or major international academic conferences. The objective of this project is to bring the research of our country in this field to reach the international advanced level and enhance the overall capacity of our country for web sensitive video surveillance.

视频分享网站、P2P、微博、播客、手机WAP网站和3GP网站的出现与发展，为敏感视频的传播提供了便利。敏感视频的传播对国家安全、社会稳定以及青少年身心健康具有危害性。本项目针对我国网络内容安全的重大需求，拟对网络敏感视频识别中一些前沿性关键问题开展研究。对于特殊类别的敏感视频识别，本项目拟重点研究基于多模态多线索概率融合的色情视频识别和结合视频画面、运动行为、音频的强度、情感表达等多种上下文信息的暴力视频识别；对于可定制样例的敏感视频识别，本项目拟重点研究基于特定目标属性学习的敏感视频识别和基于多特征语义关联的样例查重的敏感视频识别。在此基础上，完成一个网络敏感视频识别原型系统。项目的主要成果要形成发明专利等自主知识产权；同时还要在国际一流期刊和主流国际学术会议上发表一批学术论文。通过本项目的研究，要使我国在网络敏感视频识别的研究上达到国际领先水平，全面提升我国网络敏感视频信息监控水平。

目前已经全面超额的完成了本项目的研究内容。本项目根据项目计划书的规定，开展对网络多媒体信息内容监控的原理与方法的研究。先后提出了基于卷积神经网络的高精度色情图像视频识别算法、基于上下文的多视角代价敏感稀疏多示例模型及其在敏感视频识别中的应用、基于多线索深度学习的暴恐视频识别技术(包括基于特征学习的小样本特定暴恐标志检测、基于可变形卷积的暴恐视频特定物体检测、基于深度迁移深度学习的特定场景识别、基于多线索级联融合的暴恐视频识别)、基于多示例多特征的网络暴力视频识别方法、负载均衡局部敏感哈希算法及其在近似重复敏感图像检测中的应用、基于上下文敏感多示例学习的恐怖图像识别算法、基于分层上下文结构的多示例多标签学习算法及在图像标注中的应用、基于卷积神经网络的敏感视频拷贝检测方法和基于深度学习的敏感视频人类年龄估计等。开发成功便携式手机特定视频检测装备、网络异质媒体监测系统、网络直播智能监控系统。项目执行期内，在 ACM Transactions、IJCV、PAMI 等 IEEE Transactions 上和 ICCV、ECCV、CVPR、IJCAI、AAAI、ACM 上发表论文论著共38篇，申请专利17 项。此外，本项目荣获 2018 度中国通信学会科学技术奖一等奖。