基于深度学习和马尔科夫逻辑网络的特殊视频识别研究

With the development of information technique, the network traffic is dominated by video. There are more and more special videos such as violence videos affected the safety of the public. As the main method to filter the special content, the content recognition algorithm is facing series severe challenges. The traditional design methods based on feature extraction and the classifier is difficult to effectively to represent the content, and the recognition accuracy rate and the efficiency are low under the condition of the massive complex videos. The fundamental reason is that semantic concept detection is not accurate and the logical inference between semantics and recognition results is lack of theoretical support. In order to solve these problems, we need to build new relation model between semantic concepts and the recognition result, and the novel core algorithm of semantic detect. For this reason, this project uses visual consistency relevance theory to extract video key-frames, and uses deep learning theory to analyze and build related semantic concepts detection model of action, scenes and objects. Finally, we introduce a new statistical relational model——Markov logic network, which fuses first-order predicate logic which has the advantage of the high level of knowledge representation and probabilistic graphical model which is fit for the inferential and statistical learning at the bottom of the data to build the relational model between semantic concepts and recognition results. The goal of the project is to build a novel video content detection system based on semantic concept in order to achieve efficient expression of massive data and improve the accuracy and efficiency of recognition.

信息技术的发展，网络流量呈现视频化趋势，视频内容混杂化使得相关暴恐视频充斥其中，影响了社会公共安全。然而，作为过滤主要手段的内容识别算法面临着十分严峻的挑战，传统的基于特征提取、分类器设计的方法在面对海量复杂视频数据时呈现出视频内容难以有效表征、识别准确率不高效率低下等问题，其根本原因是语义概念检测不准，语义与识别结果间逻辑推理没有理论支撑等缺陷，需要构建新的语义概念与识别结果的关联关系模型及语义检测核心算法。为此，本项目将采用视觉关联一致性理论提取视频关键帧，并利用深度学习理论分析和建立动作、场景和物体相关语义概念检测模型，最后引入融合了适合高层知识表达和推理的一阶谓词逻辑和擅长底层数据统计学习概率图模型的一种新的统计关系模型—马尔科夫逻辑网络，建立语义概念和识别结果的逻辑推理模型。本项目目标为构建新的基于语义概念的视频内容检测系统，从而实现海量视频数据的有效表达，提高识别准确率和效率。

信息技术的发展，网络流量呈现视频化趋势，视频内容混杂化使得相关暴恐视频充斥其中，影响了社会公共安全。本项目针对传统的基于特征提取、分类器设计的方法在面对海量复杂视频数据时呈现出视频内容难以有效表征、识别准确率不高效率低下等问题，整理了特殊视频数据集，构建新的语义概念与识别结果的关联关系模型及语义检测核心算法。并且采用视觉关联一致性理论提取视频关键帧，并利用深度学习理论分析和建立动作、场景和物体相关语义概念检测模型，最后引入融合了适合高层知识表达和推理的一阶谓词逻辑和擅长底层数据统计学习概率图模型的一种新的统计关系模型—马尔科夫逻辑网络，建立语义概念和识别结果的逻辑推理模型。本项目应用深度学习理论框架构建新的基于语义概念的视频内容检测系统，从而实现海量视频数据的有效表达，算法有效的提高识别准确率和效率。本项目共发表和录用论文15篇，申请专利5项，实现了预期目标。