面向高层语义的跨媒体特征表达机制研究

In current Internet era, cross-media data play an increasingly important role in the information dissemination. On the whole cross-media data is multi-modality, heterogeneous, high noise and weak annotations, which brings the great challenges to the semantic analysis of cross-media. This project explores the efficient theory and solution for the unsupervised image feature representation learning, image multi-label annotation, multi-view subspace learning and cross-modal transfer learning. We discuss the following problems, the complementation and fusion between the deep model and the manifold learning model, the internal correlation among multi-labels of the images, the sharing subspace learning between the low-level visual feature and the high-level semantic description, the cross-media transfer learning mechanism with local structure preservation. Overcoming the new challenges of cross-media data, this project aims to construct the key mechanism of cross-media feature representation learning for the semantic, which can provide the strong theoretical and technical support for the practical application of cross-media intelligence computing related fields.

网络跨媒体数据在互联网时代的信息传播中发挥着越来越重要的作用。跨媒体数据总体上呈现出多模态、异构性、高噪声、弱标注等特点，这给对其进行高层语义分析带来了巨大的挑战。本项目探索有效解决无监督图像特征表示学习、图像多标签标注、多视角子空间学习、跨模态迁移学习等理论方法，研究深度模型与流形学习的互补融合、图像多标签之间的内在关联关系、底层视觉和高层语义的共同子空间嵌入机理、局部结构保持跨模态迁移机制等关键问题，克服跨媒体数据的新挑战，建立面向高层语义的跨媒体异质数据特征表达机制，为跨媒体智能计算领域的实际应用提供有力的理论与技术支撑。

互联网已成为人们工作生活的必需品，网络媒体以图像、视频、文本等多模态形式来表达内容与刻画价值思想。这些数据具有高噪声，并且各模态表征转换困难，这给对其进行高层语义分析带来了巨大的挑战。本项目探索有效解决面向高层语义的跨媒体特征表达学习机制的理论方法，研究跨模态多样性表示学习、图像多标签之间的内在关联关系、底层视觉和高层语义的空间嵌入机理、跨模态迁移机制等关键问题，克服跨媒体数据的新挑战，建立面向高层语义的跨媒体异质数据特征表达机制，为跨媒体内容理解和网络智能服务提供技术支撑。