Research On Remote Sensing Image Registration Algorithms Based On Local Features

Remote sensing technology is of great guiding significance to boost China's economic development and ensure production and living safety.As the crucial reference data of remote sensing application,remote sensing images have been widely used in military navigation,resource survey,environmental monitoring,target recognition and many other fields.Remote sensing images of a specific scene captured by different sensors will present different and complementary ground objects information,therefore,dissimilar information could be collected by observing different remote sensing images.Image registration technology is often applied to preprocess the obtained remote sensing images for the sake of eliminating the errors between images as well as preserving the various ground objects information as much as possible.Algorithms based on the local features of images are extensively used in image registration,and the local features are very robust in different images.For this reason,this thesis studies the registration algorithms for remote sensing images based on the local features.In the research of algorithms for remote sensing image registration based on key points,it was found that the traditional descriptors constructed by gradient magnitude could not meet the requirements of remote sensing image registration because of the irregular gradient variation in remote sensing images.To solve the problem,a new descriptor named Normalizing Gradient Magnitude-based Descriptors(NGMD)is proposed in this thesis,which is based on the normalization reset strategy of gradient magnitude.In the experimental section,NGMD was compared with other six descriptors based on gradient information,whose results show that the registration algorithm using NGMD has the highest accuracy.Through the study of image gradient information,gradient magnitude(GM)and gradient occurrence(GO)were found to be complementary in describing image features,based on which descriptor named Incrementing Gradient Magnitude onto Gradient Occurrence-based descriptor(IMOG)combining these two gradient information is also proposed in this thesis.The results of comparative experiments with other six descriptors,demonstrate that the registration algorithm adopting IMOG not only has the highest recall rate but also can identify more correctly matched pairs.In the corner-based remote sensing image registration algorithms,it was found that remote sensing images,photographing ground from a long distance,covers a large amount of information about ground objects(such as buildings,roads,etc.)that include corners and edges.The information such as corners and edges stored in the remote sensing images is still robust despite that fact that a series of non-linear grayscale variations occur.Based on the Distribution of Edge Pixels Along Contour Descriptor(DEPAC),a new Contour-based Corner Descriptor(CCD)is proposed in this thesis,thereby better utilizing the corners and edges in the remote sensing images.Experiments show that CCD outperforms DEPAC in accuracy.