高维结构性稀疏特征选择与图像语义理解机制研究

With the booming of social media and mobile internet applications, we are witnessing an explosion of large-scale web images. These real world images have the characteristics of high-dimensionality of heterogeneous visual features, weakly tagging, and cross-domain data source, which turns out a great challenge in image semantic understanding. Based on the recent development and research focuses in statistical analysis, machine learning, and computer vision, in this proposal, we target to develop the framework of image semantic understanding by high-dimensional structural sparse feature selection. Four key issues are explored in this proposal: utilization of structural complementary properties and adaptive selection algorithms for high-dimensional and heterogeneous image features, multi-task feature selection for cross-domain image data, prediction of high-level structural semantics for images, and analysis of consistency and stability in algorithms of structural sparse feature selection. Based on the research results of this proposal and related technologies, we will open a benchmark image dataset of structural semantic understanding and release a portal of image semantic understanding, which can be taken as a prototype to demonstrate the effectiveness of the proposed algorithms and framework. The research results of this proposal will mainly contribute to the theoretical analysis and applications of image recognition and multimedia retrieval.

社会媒体站点的兴起和移动互联应用的飞速发展使得图像数据在互联网上海量涌现，其所具有的高维、异构、弱标注和跨数据源等特性给图像语义理解带来了巨大挑战。本课题将结合统计分析、机器学习和计算机视觉等交叉领域的最新进展和热点研究，建立基于高维结构性稀疏特征选择的图像语义理解框架，主要内容包括：图像高维异构特征的结构性互补特性利用与自适应选择、跨域异构数据源的多任务特征选择、图像高层结构性语义预测和一致性特征选择及其稳定性分析。通过集成相关研究成果与技术，开放结构性图像语义理解数据集，并发布以高维结构性稀疏特征选择为基础的图像语义理解原型系统，以验证算法和框架的有效性。相关研究成果在图像识别和多媒体检索等方面具有理论价值和应用意义。

社会媒体站点的兴起和移动互联应用的飞速发展使得图像数据在互联网上海量涌现，其所具有的高维、异构、弱标注和跨数据源等特性给图像语义理解带来了巨大挑战。本项目对基于高维结构性稀疏特征选择的图像语义理解进行了研究，主要内容包括：图像高维异构特征的结构性互补特性利用与自适应选择、跨域异构数据源的多任务特征选择、图像高层结构性语义预测和一致性特征选择及其稳定性分析，并构建了结构性图像语义理解数据集和基于层次语义的图像检索原型系统，验证了算法和框架的有效性。项目成果在国内外重要期刊/会议发表论文36篇，包括IEEE Trans.论文8篇；CCF-A类会议论文7篇，CCF-B类会议论文3篇（Oral或Full Paper）；相关技术成果申请国家发明专利3项；相关算法在THUMOS 2015等国内外评测中取得优异成绩。本项目相关研究成果在图像识别和多媒体检索等方面具有理论价值和应用意义。