基于特征组对策略的分层匹配方法研究

Feature matching is the common key technique in the fields of image processing and computer vision, and the classic descriptor-based methods face the following bottlenecks: (1)Image deformation makes errors among descriptors of one single feature and damages their distinctiveness. (2)Repeated texture makes different features have similar descriptors and results in matching ambiguity. Using geometric information is an effective way for resolving the above problems, however, existing geometric optimization-based matching methods are so complicated that they are difficult to achieve. Focus on the above two bottleneck matching problems,in this project an original layered matching method is developed and investigated based on feature-pairing strategy, the main content of which includes: (1)Feature-pairing based feature description, which mainly reseaches on how to describe features using feature-pairing strategy for resolving shortcomings of traditional descriptors; (2)Feature-pairing based geometric guidance and constraint, which mainly focus on the specific form of geometric guidance and constraint from the estiblished matches to the following matches; (3)Matching algorithm designing, which mainly reasearch on how to design the algorithm under condition of image deformation and repeated textures; (4) Algorithm verification and its applications,which mainly reseach on algorithm testing and optimization, and its applications in selected typical scenes. This reseach will enrich and develop the theory and methods related to feature matching, and will be able to result in algorithms of greater performance in this field.

特征匹配是图像处理与计算机视觉领域的共性关键技术，经典基于描述子的匹配方法主要面临如下瓶颈问题：(1)图像形变导致同一特征的描述子间产生误差而影响描述子分辨力；(2)重复纹理使不同特征具有相近描述子而造成匹配歧义性。利用几何位置信息是解决上述问题的有效思路，但已有整体几何优化的方法十分复杂而不易实现。本课题将针对上述瓶颈问题提出并研究一种基于特征组对策略的分层匹配方法，主要内容包括： (1)基于特征组对的特征描述，主要研究如何通过特征组对策略进行特征描述以克服传统描述子的缺点；(2)基于特征组对的几何引导约束，明确已确立匹配对后续匹配的几何引导约束形式；(3)匹配算法设计，主要研究图像形变与重复纹理下匹配算法的设计问题；(4)算法验证及应用，主要在图像库上进行算法测试、优化并选取典型场景进行应用研究。本课题研究是对特征匹配理论与方法的丰富与发展，将有力推动这一领域产生更优算法。

特征匹配是图像处理与计算机视觉领域的共性关键技术，经典基于描述子的匹配方法主要面临如下瓶颈问题：(1)图像形变导致同一特征的描述子间产生误差而影响描述子分辨力；(2)重复纹理使不同特征具有相近描述子而造成匹配歧义性。利用几何位置信息是解决上述问题的有效思路，但已有整体几何优化的方法十分复杂而不易实现。本课题将针对上述瓶颈问题提出并研究基于特征组对策略的分层匹配方法，主要内容包括： (1) 图像特征检测，主要提出了图像中点、线、图形等特征的自动提取算法；(2)基于特征组对的特征描述，主要提出了通过特征组对策略进行特征描述的方法；(3) 特征匹配算法研究，设计了利用已确立匹配对后续匹配进行几何引导约束的分层匹配算法； (4) 算法验证及应用，主要在图像库上进行算法测试、优化并选取典型场景进行应用研究。本项目相关成果获得发明专利授权7项；发表或录用SCI收录期刊论文11篇、EI期刊论文2篇、会议论文3篇。本课题研究是对特征匹配理论与方法的丰富与发展，将有力推动这一领域产生更优算法。