【作者】 吴俊；

【导师】 张钹；

【作者基本信息】 清华大学 ， 计算机科学与技术， 2008， 博士

【摘要】 在视像概念检测中可以发现,对于同一语义概念而言,其视觉特征的潜在分布通常会随着时间发生变化。针对这种现象,本文将主要研究两个核心问题:第一,不同语义概念在不同条件下随着时间到底有什么变化规律;第二,如何在有限规模训练样本的条件下动态地更新概念模型。论文工作主要包括:1)提出了一套基于有限混合模型的、衡量在线数据潜在分布变化的度量方法,专门研究不同语义概念的潜在分布随时间变化的规律。该度量方法为在线数据流中某语义概念是否发生潜在分布的变化提供了定量的判断依据,也为建立整个在线学习系统提供了较为合理的先验知识。2)提出了监督学习条件下多时间粒度的自适应在线学习算法。该算法主要利用不同语义概念在不同时间粒度上的偏向性,并基于多时间粒度的分类器进行融合。本文着重比较了增量式和平坦式两大类分类器的选择策略,并深入研究了上述度量指标对于该在线学习算法的指导意义。当在线数据流中某语义概念的分布变化相对较快时,可以应用该在线学习算法。3)提出了半监督学习条件下在线优化的增量学习算法。该算法尽可能充分地挖掘大量存在的无类别标签的样本信息,先通过本地自适应步骤得到当前最新的本地概念模型,然后利用最新本地模型动态地更新初始的全局概念模型。该算法在一定条件下解决了半监督学习环境下的模型更新问题。当在线数据流中某语义概念的分布变化相对较慢时,可以应用该在线学习算法。4)实验结果表明,本文提出的这套基于有限混合模型的度量方法,对于在线数据流的语义概念分布特性的描述是有效的,可以为在线学习算法提供相对合理的参考信息,具有一定的普遍意义。基于体育视像和大规模TRECVID视像数据集的实验结果表明,本文提出的两种在线学习算法比同类算法更有效。更多还原

【Abstract】 In semantic concept detection of the online video streams, the underlying data distribution for a certain semantic concept in the visual feature space generally evolves over time. This thesis will tackle two key issues: i) what are the rules of the evolving underlying data distribution for different semantic concepts at different conditions? ii) how to update the concept models for the limited training samples from the current video sequence?The major contributions of this thesis comprise:1) Based on the Finite Mixture Models (FMMs), a couple of tracking measures are proposed to describe statistical properties of the evolving underlying data distribution in a quantitative way. On one hand, they can be utilized to investigate the evolving rules of different semantic concepts. On the other hand, they can provide much reasonable prior knowledge of the online data streams for the establishment of the whole online learning system.2) The Multi-granularity Adaptive (MGA) online learning algorithm in supervised learning is proposed. It mainly focuses on studying the statistical properties of the diverse granularity in the time domain for the different semantic concepts, and the corresponding classifiers fusion techniques. Two types of the classifier selection, the incremental version and the flat version, are both investigated in detail. In addition, the relationship between the enhancing capacity of the overall performance and the above defined tracking measures are also covered. This MGA algorithm is suitable for the situation where the current target concept evolves relatively quickly over time.3) The Online-optimized Incremental Learning (OOIL) algorithm in semi-supervised learning is proposed. It manages to sufficiently utilizes the statistical characteristics of the easily-collected unlabeled data samples from the newly-upcoming online data streams. The Local Adaptation (LA) step can derives the latest local concept models, which can also be used to dynamically update the original global concept models, resulting in the Global Model Incremental Updating (GMIU) step. Therefore, this algorithm has solved the problem of model updating in semi-supervised learning under appropriate conditions. This OOIL algorithm is suitable for the circumstance where the current target concept evolves relatively slowly over time.4) The experimental results show that, the proposed FMM-based tracking measures are very useful and practical to describe the evolving process of the underlying data distribution, and they are also able to be applied to the online learning applications in other domains. These tracking measures can effectively derive the evolving rule of the target concept, and provide much reasonable reference information for the above two types of the online learning algorithms. Furthermore, the experimental results based on the sports video and the large-scale TRECVID data collections demonstrate that, compared with the existing strategies, the two types of the online learning algorithms (MGA and OOIL) proposed in this thesis are more effective.更多还原