基于上下文感知的不良影像分类

项目名称： 基于上下文感知的不良影像分类

项目编号： No.61201291

项目类型： 青年科学基金项目

立项/批准年度： 2013

项目学科： 电子学与信息系统

项目作者： 田春娜

作者单位： 西安电子科技大学

项目金额： 26万元

中文摘要： 本项目主要研究多媒体通信中淫秽色情不良影像的过滤方法。不良影像内容复杂、语义信息丰富，基于低层特征的检测方法不足以描述其内容，从而易导致误判。本项目从实际应用中抽象出科学问题，旨在借鉴人类的认知机理，结合最新的视觉信息认知计算方法，根据统计推理模型提取图像隐含的语义主题，结合概率图理论充分利用不良影像中共生的上下文信息，建立新的计算模型和方法，从语义角度判定不良信息。主要研究内容包括：(1)图像隐含主题语义认知与解析；(2)基于鉴别性形变模型的人体敏感器官检测；(3)结合语义上下文的不良影像多线索集成判决。本项目涉及认知计算和统计学习的最新理论，需要针对不良影像的特点，从新的角度进行研究。研究内容具有重要的理论意义和广阔的应用前景。本项目预期在理论上有所突破，为多媒体影像语义建模、内容分析与理解方面的研究奠定理论和技术基础。

中文关键词： 语义主题解析；上下文信息；颜色显著性；鉴别性特征提取；不良信息过滤

英文摘要： This project focuses on study the pornographic filtering methods in multimedia communication. The pornographic images and videos contain complex content and rich semantic information. They are hard to be described by the low-level features, which will lead to false detection. This project abstracts scientific issues from the practical applications. We aim to obtain the latent semantic topics of images statistically by learning from human cognitive mechanism and combining visual cognitive calculation methods. Then, we use the probabilistic graphic theory to build new computational models to realize semantic pornographic filtering, which takes advantages of the co-concurrent context of the pornographic images and videos. The research content includes: (1) Latent topic cognition and analysis of images, (2) Erotogenic-part detection based on discriminative deformable model, (3) Pornographic filtering by multi-clue ensemble of semantic context. This project involves the latest theories of cognitive computing and statistical learning. We study this task through a new perspective of analysing the features of pornographic images and videos. The research has important scientific meaning and wide potential applications. During this study,we expect the breakthrough in theory, which will provide the theoretical and technica

英文关键词： semantic topic analysis and understanding；context information；color saliency；discriminative feature extraction；pornography filtering