随机映射框架下的图像语义分析与提取技术研究

项目名称： 随机映射框架下的图像语义分析与提取技术研究

项目编号： No.61501515

项目类型： 青年科学基金项目

立项/批准年度： 2016

项目学科： 无线电电子学、电信技术

项目作者： 张瑞杰

作者单位： 中国人民解放军战略支援部队信息工程大学

项目金额： 19万元

中文摘要： 图像语义分析与提取对大数据环境下的图像语义检索、智能视频监控以及机器人导航等应用具有重要的理论价值和现实意义。视觉词典模型是当前图像语义分析的主流方法，但现有方法在构建视觉词典时存在生成效率低、稳健性差以及视觉单词的同义性和歧义性等问题。基于此，本项目拟突破现有的视觉词典构建框架，研究随机映射框架下的图像语义分析与提取技术，重点研究随机映射框架下视觉词典的构建、优化和视觉词汇特征提取三个关键问题。首先，对随机映射进行拓展和改进，基于随机映射框架构建区分性和稳健性更强的视觉词典；其次，通过设计语义相似性度量准则，挖掘视觉单词间的潜在语义关联，压缩优化初始视觉词典；最后，通过构造视觉单词语义共生矩阵，挖掘利用视觉单词的空间上下文信息，提取更准确有效的视觉词汇特征。通过对上述问题的研究，有望获得更稳健高效的图像语义提取方法，为大数据环境下的图像语义检索等应用提供理论支撑和技术支持。

中文关键词： 图像语义分析；随机映射；精确欧氏位置敏感哈希；语义相似性度量；语义共生矩阵

英文摘要： The research of image semantic analysis and extraction is of great theoretical value and practical significance for image semantic retrieval, intelligent video surveillance and robot navigation. Currently, Bag of Visual Word is the state-of-the-art algorithm in image semantic analysis domain. However, the existing methods suffer from disadvantages of low generation efficiency, poor robustness and visual words' synonymy and polysemy. To solve these problems, this project aims to break through the framework of current BoVW construction and research on randomized mapping based image semantic analysis and extraction technology. Especially, this project will focus on three key problems, that is randomized mapping based BoVW construction, optimization and visual vocabulary features extraction. Firstly, expand and improve randomized mapping algorithm, meanwhile constructing more distinctive and robust BoVW based on it. Secondly, through design the proper semantic similarity measure, mine and employ latent semantic association among different visual words, thus to compress and optimize the original BoVW. Finally, through construct semantic co-occurrence matrix, mine and employ the spatial and contextual information among different visual words, thus to extract more correct and efficient visual vocabulary features. Research of the project is expected to obtain more robust and efficient image semantic extraction methods, which can provide theoretical and practical supports for image semantic retrieval and other applications in big data environment.

英文关键词： Image Semantic Analysis;Randomized Mapping;Exact Euclidean Locality Sensitive Hashing;Semantic Similarity Measure;Semantic Co-occurrence Matrix