面向遥感影像匹配的特征点检测算子性能评估

Performance Evaluation of Interest Point Detectors for Remote Sensing Image Matching

在基于特征点的匹配方法中,特征点检测是非常关键的步骤,直接影响到匹配的效果.为了确立遥感影像匹配过程中特征点算子的选择依据,本文从光谱、时相和尺度(分辨率)3个方面,选择不同类型的遥感影像作为实验数据,以特征点重复率作为评估标准,对当前主流的Harris-Laplace、Hessian-Laplace、Do G和M_SER 4种特征点检测算子进行性能评估,并分析了每一种算子的优缺点和适用范围.实验结果表明:在光谱和时相方面,Hessian-Laplace的平均重复率达到40%,性能最好,其次为Harris-Laplace和Do G,而M_SER的性能相对较弱;而对于尺度方面,M_SER表现出最好的性能,平均重复率达到35%,其次为Hessian-Laplace,而Harris-Laplace和Do G的性能较弱.

Interest point detection is a crucial step for the image matching because it directly influences the matching results. In order to establish the criterion of selecting the interest point detectors for remote sensing image matching, different remote sensing images in terms of spetrum, time and scale were selected to evaluate the four famous interest point detectors including Harris-Laplace, Hessian-Laplace, Difference of Gaussian （DoG） and Maximally Stable Extremal Regions （MSER）, and the repeatability was used as the evaluation criterion. The merits, the demerits and application of these detectors were also discussed. The experimental results show that for spectrum and time variation, the repeatability of Hessian-Laplace detector achieves 40% whichperforms best, followed by Harris-Laplace and DoG, whereas MSER performs worst. For image scale changes, MSER outperforms other detectors and its repeatability is 35% , followed by Hession-Laplace, whereas Harris-Laplace and DoG perform worse than other detectors.