Human Perception & Intelligence in Multimedia Computing

多媒体计算中的人类感知与智能人类的智能一直在推动着技术的进步。目前的研究主要集中在以机器可执行的形式模拟人类智能。机器学习方法展现了前景，但由于大量的训练数据和漫长的训练时间，也带来了挑战。为了解决这些问题，有必要在机器学习步骤之前使用鲁棒的计算机视觉和信号处理技术来优化训练特征。否则，在机器训练过程中，噪声和不相关的输入将导致计算时间呈指数级增长，甚至不收敛。我的研究目标是去除冗余数据，识别主导特征，并引入无监督算法来提供智能多媒体计算结果。选择多媒体中的主要特征，如图像、视频、信号和图形，需要在空间和时间域中进行数据和模式分析（从一维到更高维度）。文献表明，结合人类感知标准（如颜色和对比敏感度）的计算模型在特征发现、数据识别和目标识别方面是有效的。为了利用快速增长的动态和复杂的多媒体数据进行研究，我将扩展我的感知驱动模型，并重点介绍多媒体质量评估，噪声识别和多模态融合的有效方法。降维技术也将被开发以优化性能。可注意差分（JND）的确定是评估输出质量和识别噪声的重要步骤。特定应用的JND阈值（或公差），例如游戏应用或医学图像分析，需要定义，以决定输出质量是否应该接受或拒绝。多媒体数据在机器中被表示为数字值，而视觉线索是基于人类视觉系统（HVS）来解释的。HVS可能看不到数字值的变化，但这种变化可以是应用程序警报。另一方面，对HVS的可见更改对于应用程序来说并不一定是重要的。我的研究结果将进一步深入了解定量计算机输出与定性视觉评估之间的对应关系，并报告对应用程序具有重要意义的影响。我在质量评估、噪声过滤和多模态融合方面的工作将有助于跟踪动态多媒体数据的变化模式。应用将包括智能城市监控，例如监视和目标跟踪，以及在灾害发生之前预测结构或地理变化，例如山体滑坡。

Human Perception and Intelligence in Multimedia Computing Human intelligence has been driving the advancement of technology. Current state-of-the-art in the research community focuses on simulating human intelligence in a machine executable form. Machine learning methods show promise but also bring challenges due to massive training data and lengthy training time. To address these issues, it is necessary to optimize the training features using robust computer vision and signal processing techniques before the machine learning step. Otherwise, noisy and irrelevant input will lead to exponential computational time or even non-convergence in the machine training process. My research goal is to remove redundant data, identify dominant features, and introduce unsupervised algorithms to deliver intelligent multimedia computing results. Selecting dominant features in multimedia, e.g., image, video, signal and graphics, requires data and pattern analysis (from 1D to higher dimensions) in the spatial and temporal domain. It has been shown in the literature that a computational model incorporating human perceptual criteria, e.g., color and contrast sensitivity, is effective in feature discovery, data discrimination and target recognition. To conduct research using the rapidly growing dynamic and complex multimedia data, I will extend my perceptually motivated models and focus on introducing efficient methods in Multimedia Quality Assessment, Noise Discrimination and Multi-Modal Fusion. Dimensionality Reduction techniques will also be developed to optimize performance. Determination of Just-Noticeable-Difference (JND) is an important step when assessing output quality and discriminating noise. Application-specific JND threshold (or tolerance), e.g., for game applications or medical image analysis, needs to be defined in order to decide whether an output quality should be accepted or rejected. Multimedia data is represented as digital values in machines, while visual cue is interpreted based on the Human Visual System (HVS). A change in digital value may not be visible by the HVS, but such change can be an application alert. On the other hand, a visible change to the HVS is not necessarily significant for an application. The result of my study will bring further insight into the correspondence between quantitative computer output and qualitative visual assessment, and report impact which is significant to an application. My work on quality assessment, noise filtering and multi-modal fusion will help track changing patterns in dynamic multimedia data. Applications will include smart city monitoring, e.g., surveillance and target tracking, and the prediction of structural or geographical changes before disasters occur, e.g., landslides.