基于近似距离序的重复模式图像匹配算法

摘要

Due to the local ambiguities of images containing repetitive-patterns, it is difficult to match feature points reliably only by comparing similarity between local descriptors even if the disparity of viewpoint is not very large. Thus, a novel representation of geometric consistency named approximate distance order is proposed according to the space distribution of feature points. Then, an object function in hybrid form is defined by combining the matching cost of local descriptor, and the matching problem is formulated as an optimization problem with one-to-one correspondence constraints. Finally, the correspondences between feature points are obtained by maximizing the given object function via the method of probabilistic relaxation. Comparative experiments applied to various images demonstrate the algorithm is an effective approach to solving the suggested problem.%由于重复模式图像局部信息的歧义性,即使在视角变化很小的情况下,仅通过比较局部描述子之间的相似性得到的匹配结果往往并不可靠.为了解决这个问题,该文根据特征点空间的分布特性,给出了一种新颖的几何相容性表示方法——近似距离序；结合利用局部描述子获得的匹配代价,定义了一种混合形式的目标函数,将匹配问题转化为一对一约束下的优化问题；最后,通过概率松弛法求解目标函数的极大值,获得特征点之间的对应关系.在不同类型图像上的比较实验表明,所提出的方法能够有效地解决重复模式图像匹配问题.