项目名称: 基于多层上下文关系的图像目标识别关键技术研究
项目编号: No.61272351
项目类型: 面上项目
立项/批准年度: 2013
项目学科: 自动化技术、计算机技术
项目作者: 张弘
作者单位: 北京航空航天大学
项目金额: 80万元
中文摘要: 本申请从人类视觉目标识别机制中的上下文特性出发,对基于特征层、目标层和场景层三层融合的多层上下文目标识别理论进行了研究,重点解决基于上下文的目标识别任务所存在的两个难点问题:图像特征的统一表达问题及上下文关系的优化建模问题。 本研究首先从图像基本特征出发,通过建立统一的图像特征描述子,实现图像特征在上下文结构下的一致表达;其次,在一致性的特征描述基础上,进一步研究人类视觉感知过程中的上下文特性,建立基于多层上下文关系的图像目标识别模型;最后,对多层上下文目标识别模型进行优化,通过设计多层协同优化算法,降低求解多层上下文模型参数的运算量,提高识别速度。 本研究为建立符合人类理解方式的智能目标识别体系和方法提供了新的思路,在视觉目标识别领域具有重要理论意义。该研究的成果可广泛应用于安全监控、人机交互、身份识别,智能导航等多个领域,推动智能化信息处理技术在生产生活中的应用和发展。
中文关键词: 上下文关系;特征提取;目标识别;多尺度;特征融合
英文摘要: This proposal presents a multi-layer context theory for object recognition which combines the feature layer, the object layer and the scene layer in the context of an image by studying the context character of the human visual system. Solutions for context structure design and image feature extraction in context environment will be provided in this research. Firstly, a unified image feature representation is presented which builds basic description of objects for context structure. Secondly, based on the unified feature description, a multi-layer context system is proposed to simulate human's perception process in recognizing objects. Finally a collaborative approach is proposed to reduce computation load and cost. This project provides new thoughts for building intelligent recognition systems, and is of great importance in the field of visual object recognition. The research can produce achievements in the applications of security surveillance, human-computer interaction, identity recognition and intelligent navigation etc. These achievements can also promote the development and application of intelligent information systems in both manufacturing industry and daily life.
英文关键词: multi-layer context;feature extraction;object recognition;multi-scale;feature combination