大尺寸高分辨率差异图像的结构化分层细分配准研究

项目名称： 大尺寸高分辨率差异图像的结构化分层细分配准研究

项目编号： No.61302171

项目类型： 青年科学基金项目

立项/批准年度： 2014

项目学科： 无线电电子学、电信技术

项目作者： 周武

作者单位： 中国科学院深圳先进技术研究院

项目金额： 24万元

中文摘要： 关于图像间复杂非线性和局部几何变形的配准，始终面临着两个重要的问题：如何匹配特征以及选择何种映射变换模型。尽管采取径向基函数，能够较好地解决配准中的映射变换问题，但复杂非线性、局部几何变形或大尺寸高分辨率图像的特征匹配仍然是没有被解决的重要难题。由于图像间复杂的变换关系和存在较大的局部差异，目前基于特征或区域的图像配准方法很难获得稳定、可靠的匹配特征。本项目拟从结构化属性分布匹配和分层细分的角度，解决大尺寸高分辨率的全局和局部差异图像间的配准问题。首先，由图像局部方向属性构建精确稳定的全局方向分布属性直方图，通过归一化方向直方图配准检测图像间旋转角度差异；然后，由图像局部显著特征的尺度属性构建全局尺度属性分布直方图，通过归一化尺度直方图配准实现图像间的尺度差异检测；最后，在图像粗调整后拟采用多层细分思路实现局部差异区域的配准，并最终高效率地实现大尺寸高分辨率全局和局部差异图像间的精细配准。

中文关键词： 图像配准；图像特征；方向分布直方图；医学影像；肝癌影像

英文摘要： When registering images with non-linear and local geometric distortions, it always exists two basic problems-how to match features and what mapping functions to use for registration. Although the second one can be solved at least on theoretically level by using appropriate radial basis functions, the first problem is generally unsolvable due to the nature of registration of images with complex non-linear, locally dependent geometric distortions,and high resoluation. Due to the large image intensity and geometric differences, it becomes infeasible to generate stable and reliable corresponding features using certain approaches which are generally based on features or regions for image registration. In this proposal, we will develop a novel approach to achieve registration of such images with large local geometric differences. In spite of considerable differences between two images of a scene, through constructing the relationship of local attributes and global attributes of images, we can tell whether the images have the same orientation, or if one image is rotated with respect to the other, or if one image has a larger/smaller scale than the other. Firstly, determining the rotational differences between images through normalized histogram matching of accumulating local dominant orientations, and then determining

英文关键词： image registration；image feature；orientation distribution histogram；medical image；liver image