一种新的快速局部不变特征算法  

New fast local invariant feature algorithm

作　　者：王灿进 孙涛[1] 陈娟[1] 

机构地区：[1]中国科学院长春光学精密机械与物理研究所激光与物质相互作用国家重点实验室,吉林长春130033 [2]中国科学院大学,北京100049

出　　处：《红外与激光工程》2014年第6期2013-2020,共8页Infrared and Laser Engineering

基　　金：长春市科技计划项目(2013270);吉林省科技发展计划项目(20126015)

摘　　要：针对传统局部不变特征算子主方向提取不准确和匹配阶段过于耗时的问题,提出一种基于RI-LBP算子和混合spill树的快速局部不变特征算法。首先提出一种FAST-Difference算法,提取出模板图像和待匹配图像的稳定特征点,然后使用旋转不变的RI-LBP描述符计算特征向量,最后对特征向量集使用混合spill树进行匹配并使用RANSAC算法剔除误匹配点。RI-LBP算子自身的旋转不变性能够在一定程度上克服特征点主方向确定不准确的缺点,使特征描述符的提取更加稳定,并生成更简单的53维局部不变特征描述符。混合spill树相对于kd-tree省略了回溯过程,对于高维数据拥有更好的匹配效率。实验证明:该算法与SURF算法描述能力相近,旋转和光照条件下比SURF性能更优,并且匹配速度更快。In order to solve the problem that traditional local invariant descriptors extracted inaccurate main direction and spent too much time in matching vectors, a new method for fast image registration based on RI-LBP algorithm and hybrid spill-tree was proposed. Firstly, stable feature points of template image and image to be matched were extracted by the proposed FAST-Difference Algorithm. Feature vectors were calculated using rotation invariant RI-LBP descriptors. At last feature vector sets were matched using hybrid spill-tree and mismatching points were eliminated by RANSAC. The problem that the main direction couldn＇t be extracted accurately was conquered because of the rotation invariant of RI-LBP , which means the feature descriptors were more stable. At the same time the feature vectors contain contained 53 dimensions, which are more simple. Spill-tree had better matching efficiency for high-dimensional data because it omitted the process of backtracking. The experiment results indicated that the proposed method cost much less time while retained nearly the same describing performance with SURF and achieved better performance in rotation and illumination changes.

关 键 词：FAST-Difference RI-LBP 混合spill树 RANSAC 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]