视觉内容动态迁移与关联理解的随机图计算及学习方法研究

In recent years, network and social characteristics of visual data has become increasingly evident. In addition to its inherent characteristics including high dimension, mass and disordered distribution, the visual data has also become associated with other heterogeneous media data in different levels of relationship due to relevant concepts, themes, and events. Most of the existing visual semantic-aware computing works dealt with a single media object in isolation, and focus only on its inherent characteristics. This often leads them to situations like many difficulties and being low efficiency in computation, and only low level semantics obtained. This is mostly due to their inadequate attention on network and social characteristics of visual data, including heterogeneous distributed and interrelated relationship. This project does research on the problem of semantic understanding of large-scale visual data, via exploiting interrelated association among heterogeneous visual data, as well as integrating with visual cortex-like saliency computing model. Through the research, we try to propose stochastic attribute graph models to represent interrelated relationship, and semantic computing techniques including feature dimension reduction, the mapping from the low-level feature to high-level semantics, and statistical learning, to overcome the curse of dimension and shorten the semantic gap. The expected results will be beneficial to breaking through the bottleneck of visual data understanding, and enriching the visual computing theory, as well as providing novel, highly efficient visual data processing technologies.

视觉数据的网络化、社会化特征日趋明显，除高维、海量等固有特性外，更与蕴含着相关语义且无序分布的异构数据以不同层次的关联关系共存于网络。现有大多数的视觉语义计算由于孤立地处理单个视觉对象且专注于其固有特性，却对视觉数据的网络异构分布及相互关联等特性重视不够，经常性地陷入求解困难、计算效率低或仅能理解低层语义等困境。本项目聚焦于网络异构视觉数据之间的关联关系，结合类生物视觉皮层的显著性计算模型，研究大规模网络视觉数据的关联推理与语义理解问题。重点研究视觉随机属性图(attribute graph)表征模型与语义动态迁移学习方法，拟提出高维异构视觉数据的关联关系随机属性图、特征降维与关联、特征到高层语义的映射关系等表征模型及语义迁移学习方法，以克服视觉数据语义感知计算中的维数灾难问题，缩短视觉数据理解的"语义鸿沟"。预期的研究成果有助于突破视觉数据语义理解技术瓶颈，发展新型视觉媒体处理技术。

在基金项目的资助下，紧紧围绕视觉内容动态迁移与关联理解的随机图计算及学习方法开展研究，在八个方面取得了重要进展。在具有不变性的视觉语义结构表征与提取方面，提出了一种基于局部线性几何结构不变的正确匹配识别算法。在视觉数据关联属性随机图表征模型与推理算法方面，提出了一种图像检索中的民主扩散式聚合算法。在网络图像语义理解的关联规则模型与学习算法方面，提出了一种借助极弱监督的视频目标发现与协同分割算法。在基于语义理解的网络图像的自主分类与收集方面，提出了一种基于上下文感知线索的显著性物体提取模型与目标识别算法，和一种基于超复数傅氏变换谱对比度显著性检测算法。在图像数据真实性鉴别技术方面，提出了一种基于视觉模型的感知图像哈希方法，进而提出了一种针对LSB替换攻击的图像隐写取证方法。在基于网络标签与视频内容分析的视频摘要技术方面，提出了一种基于形状特征的运动目标分割方法及视频摘要方法。在系统整合方面，建立了一个Web图像/视频数据自主采集、筛选与标注系统，实现了对Web类别数据的自动下载、筛选和标注。在基准数据库建立方面，首次分别建立并标注了一个包含噪声的Web类别图像数据库和一个包含噪声的Web类别视频数据库，为本课题研究和同类研究提供了基准数据。目前课题已在国际权威期刊和重要国际会议上发表论文31篇，包括IEEE Trans. on Pattern Analysis and Machine Intelligence、IEEE Trans. on Information Forensics and Security、IEEE Trans. on Multimedia、IEEE Trans. on Intelligent Transportation Systems、Pattern Recognition和Computer Vision and Image Understanding等重要国际期刊发表论文17篇，ACM Multimedia、ICMR等重要国际会议发表论文14篇。获授权国家发明专利4项。在项目执行期间，项目负责人薛建儒老师被评选为长江学者。培养博士生7人，硕士生13人。