基于区域重叠核加权Hu矩的SIFT误匹配点剔除算法  

SIFT mismatching points eliminating algorithm based on region overlapping kernel weighted Hu moment

作　　者：宋丹[1] 唐林波[1] 赵保军[1] 

机构地区：[1]北京理工大学信息与电子学院,北京100081

出　　处：《系统工程与电子技术》2013年第4期870-875,共6页Systems Engineering and Electronics

摘　　要：针对尺度不变特征变换(scale invariant feature transform,SIFT)算法在特征点匹配时容易出现误匹配现象,提出了一种基于区域重叠核加权Hu矩的SIFT误匹配点剔除算法。该算法首先通过对SIFT描述子区域内的重叠4邻域计算Hu矩,生成能够描述纹理特征与轮廓特征的种子点描述子;其次,根据描述子的区域特点利用核函数对种子点描述子进行加权,生成63维区域重叠核加权Hu矩描述子;最后用巴氏(Bhattacharyya)系数计算归一化后描述子的相似度,并剔除相似度较小的匹配点。将该算法与其他3种算法进行对比,实验结果表明,该算法的鲁棒性最强,实时性较高,综合性能最优。A kind of scale invariant feature transform （SIFT） mismatching points eliminating algorithm based on region overlapping kernel weighted Hu moment is proposed to solve the mismatching problem of SIFT. Firstly, this algorithm computes Hu moment overlap in the 4-neighborhood of SIFT descriptors, which can generate the seed point descriptor with contour feature and texture feature. Then according to the characteristic of the SIFT region, this algorithm uses a kernel function to weighting the seed point to generate a 63-dimensional feature point descriptor. Finally, the Bhattacharyya coefficient is used to compute similarity of matching points, and eliminates those matching points with low similarity. The proposed algorithm is compared with three other algorithms, the experimental results show that the proposed algorithm has the best robustness and real-time feature, and a good comprehensive performance also.

关 键 词：误匹配点 HU矩 区域重叠 核加权 巴氏系数 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]