基于协同学的并行多层次反馈图像理解研究

项目名称： 基于协同学的并行多层次反馈图像理解研究

项目编号： No.60875012

项目类型： 面上项目

立项/批准年度： 2009

项目学科： 生物科学

项目作者： 高隽

作者单位： 合肥工业大学

项目金额： 30万元

中文摘要： 本课题在协同模式识别方法研究基础上，为解决视觉原型向量描述，从场景与目标的视觉特征描述入手，研究了颜色形状信息、全局信息、兴趣点信息、稀疏基元信息对视觉目标和场景识别的影响，同时借LabelMe工具对场景图像的语义标记分析，并使用WordNet词汇对场景词汇进行编码；针对视觉特征的序参量重构过程，讨论了基于AP聚类的原型向量采样方法，并对原始特征利用PCA、ISOMAP全局流形与LLE、LTSA局部流形进行数据维数分析；针对协同演化势函数方程中的协同竞争项参数估计，利用已有分类方法定义的判决空间，对歧义数据点的近邻样本进行分析，完成数据的动力学演化过程；利用字典学习策略完成图像理解中场景与目标编组信息，讨论视觉显著性约束下的SIFT特征协同识别过程，讨论GIST特征下的场景/目标双层协同模式识别过程。

中文关键词： 图像理解;协同学;场景描述;场景中的目标识别

英文摘要： This subject is based on the collaborative pattern recognition, in order to describe the visual prototype vectors, we do research on the visual features of scenes and objects. We have studied the information of color, shape, global, points of interest, sparse coding bases and their impact of pattern recognition. With the use of LabelMe tools, we can achieve the semantic labelings of scene images and WordNet to encode scene vocabularies. In order to obtain reconstructed order parameters for the visual features, we have discussed the prototype vector sampling method based on AP cluster and also have used global manifold methods such as PCA ,ISOMAP and local manifold methods such as LLE, LTSA to reduce the dimensions of original features, and then decision space can be defined by existing classification methods to analyze the neighbor samples of ambiguous data points to complete the dynamical evolution of data. The finally discussion about the SIFT features collaborative pattern recognition process has been investigated with the constraint of visual significant and the similar issues about the scene / object collaborative pattern recognition process using GIST features are analogous.

英文关键词： Image understanding; Synergetics theory; Scene description; Object recognition in scene