选择性注意驱动的图像语义理解方法与计算模型研究

项目名称： 选择性注意驱动的图像语义理解方法与计算模型研究

项目编号： No.90820003

项目类型： 专项基金项目

立项/批准年度： 2009

项目学科： 金属学与金属工艺

项目作者： 黄铁军

作者单位： 北京大学

项目金额： 50万元

中文摘要： 借鉴人类视觉系统中稀疏编码、成组编码、选择性注意等生理心理机制，提出并研究一种新的图像语义表示和计算模型，该模型以视觉特征相对稳定且与客观事物具有明确映射关系视觉单元(Visual Units)为语义表示、分析的核心，通过对规模化的图像和关联信息的非监督学习实现对图像的语义分析与理解。在图像不变特征提取、图像基元提取、要素图生成等图像处理和计算机视觉已有成果基础上，研究利用选择性注意模型，实现视觉单元的自动抽取，利用视纹实现视觉单元的比对和识别，采用优化学习的方法自动构建视觉单元词典，用内容关联分析方法实现图像和视觉单元的自动文本标注，从而建立一套较为完整的图像语义理解计算模型、方法和技术体系，在特定场景监控图像的自动解释与索引系统、海量互联网图像语义分析与标注两个方面进行实验验证。

中文关键词： 图像理解;选择性注意;视觉单元;视觉指纹;语义分析

英文摘要： According to physiological and psychological mechanisms such as sparse coding, group coding and selective attention of human vision system (HVS), the project proposes and studies a new image semantic representation and computing model. By employing visual unit which has stable visual features and can be mapped to objects in physical world directly as the core for semantic representation, image understanding will be carried out by unsupervised learning on the relationship and associated data of amount of images. Based on previous work on image processing and computer vision such as extraction of the variant features, visual primitives and primal sketch, visual units are extracted under selective attention model and compared and identified by visual fingerprinting technology. Then, a visual unit dictionary is generated from the visual units above extracted with textual description being annotating automatically by relational analysis of images and their surrounding texts. As a result, a set of models, approaches and technologies of image semantic understanding will be established and will be tested in two experiments: the surveillance scene automatic interpreting and indexing and the Internet image semantic analysis and annotation.

英文关键词： Image understanding;Selective attention;Visual unit;Visual fingerprinting;Semantic analysis