Research On Remote Sensing Image Registration Algorithms Based On Feature Points And Applications

Image registration is one of the fundamental tasks in the field of image processing, and it has been wildly applied in computer vision, pattern recognition, medicine, military and remote sensing. Precisely speaking, image registration is the process of overlaying images (two or more) of the same scene, taken at different times, from different viewpoints, and/or by different sensors.Image registration technique is the crucial step in remote sensing image application field. During the last few decades, many researchers have done a lot of research work, and have proposed a large number of methods to implement image registration. As the rapid development of sensor, aerospace and space technology, the types of remote sensing images and the differences among them increase gradually. Different users have their own special demands for image registration, so there are still many problems and difficulties remained to be improved and solved. In this dissertation, the theory and methods of image registration have been deeply and systematically studied. This dissertation aims at improving the speed, accuracy and automation of image registration algorithm. According to the characteristics of nephogram and large multi-spectral remote sensing images, novel registration methods are proposed. The primary work and remarks of this dissertation are as follows:(1) A hierarchical transformation frame for nephogram is proposed. The reasons of the deformations for the serial nephograms have been deeply analyzed from both internal and external aspects. It can be concluded that serial images obtained by meteorological satellite have both global rigid deformation and local nonrigid deformation. Due to this characteristic, a hierarchical transformation frame is suggested. The global rigid deformation of the image is mainly caused by the position changes of the sensor, and it includes rotation and shift. Cloud’s distortion results in the local nonrigid deformation. In the hierarchical transformation frame, the global rigid deformation is corrected by rigid registration method at first, and then the local nonrigid deformation should be adjusted by proper nonrigid registration method. The hierarchical transformation frame has been applied to the fully automated registration system of nephograms, and the results indicate that the proposed hierarchical transformation frame can not only obtain satisfied results, but also decrease the runtime of the process.(2) An improved rigid registration method based on Forstner feature points is proposed to register the global rigid deformation of the serial nephograms. Forstner operator can be simply calculated and has high precision. Therefore, Forstner operator is used to extract feature points. Feature matching contains two steps:coarse matching step and fine matching step. The coarse matching step is achieved by Normalized Cross-Correlation (NCC). Through setting up the search window, the range of the macth points is shrinked. Meanwhile, the efficiency and accuracy of coarse matching step are increased. An improved threshold-adaptive space distance constraint is proposed in the fine matching step, which enhances the adaptability and robustness of the algorithm to different types of images. The experimental results clearly indicate that the proposed approach can realize image rigid registration automatically, efficiently and is with sub-pixel precision.(3) Nephograms exist some local areas with nonrigid deformation. Therefore, an improved nonrigid registration method based on B-spline is proposed. First, control grid with fixed grid spacing is structured on the nephogram, and the B-spline curve function is used to simulate the gray deformation field of the image. According to the B-spline theory, the position change of the grid control vertexes results in the corresponding change of the gray deformation field of the image. So image registration is realized by disturbing the position of the grid control vertexes. The proposed method applies the local update strategy based on the theory of greedy algorithm instead of the original whole update strategy. The greedy local update strategy improves the nonrigid deformation control ability of the algorithm. Meanwhile, it reduces the iteration times of the algorithm. The experimental results show that the improved nonrigid registration method based on B-spline can effectively correct the local nonrigid deformation in nephogram, and the efficiency of the algorithm is improved significantly.(4) The large multi-spectral remote sensing images demand fast full-automatic registration process. Aiming at the above requirements, a gridding registration method based on SIFT (Scale Invariant Feature Transform) feature points is proposed. The gridding theory is introduced in the algorithm. According to the characteristic of large multi-spectral remote sensing images, the feature grids extraction criterion is proposed. First of all, a two-degree regular mesh is formed on the image. Then, the feature grids are chosen by grey value, entropy and feature distribution uniformity rule. Pixels in these feature grids instead of all the pixels in the image participate in subsequent processing steps, so the amount of calculation gets decreased. The selection of feature grids provides the foundation for the parallel computing of feature extraction step and feature matching step. For each feature gird, the SIFT threshold is automatically determined by the corresponding information entropy of the feature grid. It reduces the artificial intervention, and makes the algorithm full automatic. Due to the characteristic of large multi-spectral images, the primary feature matching step in SIFT is improved. According to the priori knowledge, the space position constraint is applied to the primary feature matching step. The space position constraint reduces the amount of calculation, and increases the accuracy rate of the primary feature matching step. The experimental results indicate that the proposed registration algorithm is fast and full automatic, and can achieve sub-pixel precision.