Study And Realization On Auto Registration Of Remote Sensing Image Based On Sift

Image registration is a fundamental problem in remote sensing, medical computer vision and many other fields, and it also is an important prerequisite for image stitching and fusion. Especially, registration accuracy can greatly affect the mapping quality in the remote sensing image processing.The essay firstly studies on some traditional methods of feature extraction, such as SUSAN, Harris etc. and analyses their strengths and weaknesses. For the disadvantages of traditional methods in threshold adaptively and rotation changes, and it could not meet the needs of registration requirements for remote sensing images of different phase, resolutions and scales. The essay introduces the SIFT feature extraction and matching algorithms and do further researches on its theory and implementation, as well as some relative experiments. For problems of error-match in the SIFT algorithm, author prompts utilizations of slope to remove the error-match points in the processing of matching as long as image rotation is not very obvious. Through the experiments, the SIFT algorithm has been improved to be applicable.Take remote sensing images for example, author realize several processing functions of image registration, stitching, fusion based on the improved SIFT algorithm for extracting feature points and matching. As the SIFT algorithm threshold values are set according to image data, this process does not require too much human intervention, which can be done automatically to thrift plentiful mental and physical resources, avoid errors caused by artificial picking and greatly elevate the efficiency and accuracy of image registration.