大数据流中低层与高层惊奇事件检测的研究

The concept of information is central to science, technology, and biological function. Shannon's theory of information, although eminently successful for developing modern computer and telecommunication technologies, does not capture subjective and semantic aspects of information that are not related to its transmission, but rather to the subjective value data may have for an observer. Consequently, Shannon's theory has enjoyed only limited applicability in domains that require extraction of subjectively more surprising, interesting, or important information from large datasets. Here we will use our recent mathematical definition of surprise within the Bayesian framework to identify surprising events as a front-end towards then finding more "important" information. Surprise as defined by Itti & Baldi (2009) quantifies how data affects a natural or artificial observer, by measuring the difference between prior and posterior beliefs of the observer. Surprise is better suited to studying subjective aspects of information than Shannon's theory, and has been shown in human eye-tracking experiments to significantly better predict what attracts the gaze of observers watching TV. The proposed research is articulated around two components: (1) A theoretical framework that develops a Bayesian theory of surprise to quantify the subjective value of data to different classes of observers, focusing on how data may change their current beliefs about the world. Starting from surprise, which is mainly concerned with initial orienting towards potentially important events, we will develop a theory to further filter out surprising but irrelevant events given a particular task or goal. This will rely on machine learning to link surprise, in a task-dependent manner, to the higher-level notion of relevance, to give rise to a compound measure of "importance". (2) Applications to video and text analysis: (a) Video summaries that focus on low-level (image-level) important events in the broad context of video surveillance. (b) Video summaries now computing surprise and importance at a higher level of tracked objects and their interactions.(c) Detecting surprising and important web pages within collections gathered from a search engine query.

信息的概念对于科学、技术和生物学功能都至关重要。香农的信息理论尽管在现代计算机的发展中起到了显著的作用，但是没能覆盖与信息传输无关但却与观察者有关的信息的主观和语义特征。因此，香农理论仅在有限的领域内具有一定的应用价值。为此，我们提出了另一种衡量数据对观察者主观作用的算法。我们首先从数学上定义“惊奇”的概念，然后通过贝叶斯框架在大规模数据流中找到“惊奇”的数据，从而进一步找到“重要”的信息。“惊奇”的基本定义是通过测量对象在事前和事后概率分布的差异，从而量化数据对于不同对象的主观价值。“惊奇”理论比香农定理更适合于对信息主观层面的研究，并且我们已经在人类眼部跟踪实验中验证了其有效性。本课题的研究将围绕两部分展开：一是建立基于贝叶斯的惊奇理论，量化主观数据的价值，借助机器学习技术从视频或文本流找出高级别的“惊奇”且相关的数据；二是在视频摘要、文本分析等应用中验证和使用该理论。

针对传统的卷积神经网络由于以固定大小图像输入而无法充分获取肺结节图像多尺度以及上下文特征信息，提出了一种基于2D多尺度深度残差网络模型实现肺结节良恶性分类，通过设计的多卷积结构，模型能够学习更多的多尺度特征，采用深度残差结构增加网络深度，提高网络学习抽象特征信息能力，同时将最后一个残差结构的特征输出与平均池化层输出进行特征融合提升模型特征丰富性；卷积神经网络的池化层通常用来扩大感受野进行数据降维操作，但其由于其的降采样操作会导致许多细节信息随着随着网络深度增加而丢失，针对这一问题，构建了基于3D空间金字塔膨胀卷积网络模型实现对肺结节的良恶性分类模型，通过采用具有不同膨胀系数的膨胀卷积，实现在不改变特征分辨率的基础上获取不同感受野下的特征信息，有效地提高肺结节良恶性分类准确度；提出了一种由粗到细的深度全卷积神经网络用于从眼底图像中提取血管，并尽量保留更多的原始图像信息用于特征提取阶段，在精优化阶段，将条件随机场作为全卷积神经网络的一部分，用以充分利用眼底图像中像素之间的空间关系，通过这种方式，将粗细两个阶段集成到一个整体的网络进行训练，以改善网络的分割效果，此外，在对图像处理中还进行了必要的预处理和后处理操作以提高图像的输入输出质量。

内容获取失败，请点击重试