基于多尺度特征的遥感图像密集匹配方法  

Method of Remote Sensing Images Dense Matching Based on Multi-Scale Features

作　　者：胡少兴[1] 王惟达 柴进[1] 张爱武[2] 

机构地区：[1]北京航空航天大学机械工程及自动化学院,北京100191 [2]首都师范大学三维信息获取与应用教育部重点实验室,北京100037

出　　处：《光学学报》2013年第B12期52-58,共7页Acta Optica Sinica

基　　金：国家科技支撑计划（2012BAH31B01）、国家自然科学基金（40601081,41071255）

摘　　要：提出一种利用多尺度特征的无人飞艇遥感平台获取的序列航拍图像生成密集匹配视差图像的方法。运用尺度不变特征变换（SIFT）算法从两幅相邻图像中提取特征点，基于欧氏距离进行粗匹配,并通过缩小搜索范围提高粗匹配效率，再通过随机抽样一致性（RANSAC）算法估计基础矩阵，利用对极几何约束关系剔除误匹配，进行精确匹配，提高匹配的稳健性和精度，然后利用区域生长算法进行特征点密集匹配生成相应视差图像。实验表明算法在保持稳健性的同时，可以降低时间复杂度，获得相当规模的密集匹配点，得到良好的可视效果。The algorithm adopting multi-scale features to generate parallax images with dense matching for image sequences collected from unmanned airship platform is proposed. Scale invariant feature transform （SIFT） algorithm is adopted to extract features from adjacent images firstly. The initial matches based on Euclidean distance are carried out, and the efficiency of feature matching is improved by reducing the search area. The random sample consensus （RANSAC） algorithm is adopted to estimate the fundamental matrix, and the epipolar geometry constrain is used to delete the wrong matches to improve the robustness and accuracy of feature matching. The region growing algorithm is adopted to carry out dense matching and produce parallax images. Experimental results indicate that the proposed algorithm keeps robustness, reduces the time complexity, generates lots of dense feature matches and gets good visual effects.

关 键 词：图像处理 尺度不变特征变换算法 随机抽样一致性算法 视差图像 特征匹配 区域生长 

分 类 号：TP274.2[自动化与计算机技术—检测技术与自动化装置]