Associative mining for intelligent organisation and analysis of multimedia information

在当今以媒体为中心的社会中产生的大量多媒体信息给最终用户、企业和研究人员带来了一系列挑战：简单地说，用于高效组织、导航和分析的技术无法跟上步伐。部分问题在于，如果建立一个“有趣”事件的先入为主的概念，并迫使媒体收集围绕这个概念进行组织，从而促进快速和相关的访问，媒体消费（浏览/分析）将会得到更好的服务。重组大型媒体集合以进行非线性访问的问题并不新鲜，它为基于内容和基于概念的检索提供了很大的动力：这两个核心研究前沿支撑着当前视频搜索的最新技术。然而，这样的成功是高度依赖于注释的，而且越来越明显的是，集合的增长速度很快超出了我们可靠注释的能力。迫切需要在没有先验知识的情况下自动组织多媒体内容的方法。这可以被视为无监督模式或事件“发现”中的问题，其中媒体流的内在表示被寻求作为其重组的基础。鉴于视觉注意建模的最新进展，该模型考虑了引导人类视觉/听觉注意的自然队列和反应，我们问：多媒体数据中是否存在某些内在模式或不规则性，可以更恰当地反映语义事件的自然映射？具体来说，拟议的研究计划将通过调查自然的、自组织的无监督学习方法与视觉注意机制之间可能的协同作用来解决事件发现问题，这些机制似乎显著地引导了人类的兴趣。这项研究的一般范围将导致工具的发展，以服务于广播新闻，体育，电视和电影制作，会议，个人生活，编年史，以媒体为中心的管理系统和PVR，安全和无人监控，生物医学图像信息学和生物识别的总结，导航和消费方面的一系列紧迫缺陷。

The overwhelming amounts of multimedia information generated in today's media-centric society present a host of challenges to end-users, businesses and researchers alike: quite simply, technologies for efficient organization, navigation and analysis cannot keep pace. Part of the problem is that consumption of media (browsing/analysis) would be far better served by building a pre-conceived notion of what constitutes an "interesting" event, and forcing the media collection to be organized around that, thereby promoting fast and relevant access. The problem of restructuring large media collections for non-linear access is not new, having provided much impetus for both content- and concept-based retrieval: the two core research fronts underpinning current state-of-the-art in video search. Such successes, however, are highly dependent on annotation, and it is becoming increasingly evident that collections are fast outgrowing our abilities to reliably annotate. Methodologies are desperately needed to automatically organize multimedia contents without the benefit of prior knowledge. This can be cast as a problem in unsupervised pattern or event "discovery", wherein an intrinsic representation of a media stream is sought as a basis for its reorganization. In light of recent advances in visual attention modeling, which considers the natural queues and responses that direct human visual/aural attention, we ask: is it possible that there exist certain intrinsic patterns or irregularities in multimedia data, that can more appropriately reflect natural mappings to semantic events? Specifically, the proposed research program will address event discovery by investigating the possible synergies between natural, self-organizing approaches to unsupervised learning and the visual attentive mechanisms that appear to significantly guide human interest. The generic scope of this research will result in the development of tools to serve a range of pressing deficiencies in the summarization, navigation and consumption of broadcast news, sports, TV and film production, meetings, personal lifelogs, eChronicles, media-centric management systems and PVR's, security and unmanned surveillance, biomedical image informatics and biometrics.