面向弱纹理空间目标的特征点匹配方法

Feature Point Matching Method for Weakly Textured Spatial Objects

特征点提取与匹配是遥感图像处理中关键的一环,目前成熟的算法大多面向对地成像类型的遥感图像,对于空间目标的遥感图像,没有考虑成像条件与探测平台的影响因素,特征点匹配质量较差。针对空间目标的匹配精度不高这一问题,文章提出了一种基于聚类的特征点匹配算法。首先,根据空间目标的重复弱纹理进行特征点提取与描述,再利用特征点的空间位置进行聚类,并对特征点簇进行匹配;之后将特征点的主方向减去目标整体方向,利用特征点主方向对每一个点簇进行再分组,并完成特征点匹配;最后利用最近邻次近邻比率方法和随机样本一致算法(RANSAC)剔除外点。采用该特征点匹配方法进行的模拟成像数据实验结果表明,对于空间目标图像,基于聚类的特征点匹配较直接匹配,匹配数量的提升最高可达50%,重投影误差优于1/4个像元。文章提出的这一方法使用目前通用的各种特征描述子,能够大幅度提高空间目标图像特征点匹配的数量与精度。

Feature point extraction and matching are crucial aspects of remote sensing image processing.Currently,most mature algorithms are designed for remote sensing images of Earth’s surface,with little consideration for the imaging conditions and the influence of the detection platform on spatial target images.As a result,the quality of feature point matching for spatial target images is often poor.To address the issue of low matching accuracy for spatial targets,this paper proposes a clustering-based feature point matching algorithm.First,feature points are extracted and described based on the repetitive weak textures of spatial targets.Then,clustering is performed using the spatial positions of the feature points,and matching is carried out for the clusters of feature points.Subsequently,the main direction of each feature point cluster is adjusted by subtracting the overall direction of the target.This adjustment is used to further group the points within each cluster,facilitating feature point matching.Finally,outliers are eliminated using the nearest neighbor-to-second-nearest-neighbor ratio method and the Random Sample Consensus algorithm(RANSAC).Simulation experiments with imaging data using this feature point matching method demonstrate that,for spatial target images,clustering-based feature point matching outperforms direct matching.The improvement in the number of matches can reach up to 50%,and the reprojection error is better than 1/4 pixel.The method proposed in this paper utilizes various commonly used feature descriptors,significantly enhancing the quantity and accuracy of feature point matching for spatial target images.