Local Feature-based Remote Sensing Image Registration

With advances and developments in remote sensing technology, we have access to a wide variety of image types. The requirements of remote sensing image fusion and change detection are increasing, the request of image registration as well. In this paper, aiming at specific applications, two registration algorithms are designed, and balance in the accuracy of registration algorithm, efficiency, reliability. According to the distortion characteristics of remote sensing images, propose two registration algorithms for different business needs to focus on accuracy and efficiency of each.First, most kinds of detectors and descriptors are described and summarized in a comprehensive way, and draw its performance, stability and binding properties and so on, to provide a theoretical basis to select the appropriate detector and descriptor for the registration algorithm.Second, aiming at the need for high precision, propose the SIFT-based remote sensing image registration algorithm. SIFT operator is used as both the detector and descriptor of SIFT-based image registration algorithm. Using epipolar geometry, which has good stability and high accuracy, removes false matching points. Then, fine-tune the matching points using correlation coefficients. The algorithm takes on high precision and adaptability.Third, aiming at the real-time requirement, propose the remote sensing image registration algorithm based on the multi-scale Harris. The registration algorithm use multi-scale Harris with lower complexity as feature detector, and use the PCA-SIFT with lower dimensions as feature descriptor. The use of two-way matching algorithm to match the characteristics points. The registration algorithm is high real-time, moreover ensuring a certain degree of accuracy and reliability.Fourth, compare and analyze experimental results of two registration algorithms. RMSE is used to describe the accuracy of two registration algorithms. Compare and analyze the complexity of two registration algorithms. Compare the adaptability of two registration algorithms to common image transformation model. Finally, come to the conclusion: two registration algorithms are complementary.