Multi-source Remote Sensing Image Registration Based On Scale Invariant Feature Transform

Aiming to the problems that the registration of remote sensing images, especially multi-source remote sensing image, can not meet the requirements of fast processing, automation and highly precision, this topic was proposed. The topic is one of the key research issues of the project (Remote-sensing Monitoring Information Service System of Beijing Region) which was subsidized by the China's Special Funds for National Technology R&D Program. The research tasks of the topic is with the purpose of extracting matching features by using of improved Scale-Invariant Features Transform (SIFT) algorithm and realizing fast accuracy registration of multi-source remote sensing image by using of optimal matching strategy. Great grayscale, perspective and resolution differences as well as noises caused by environmental factors among multi-source remote sensing images increase the difficulty of registration. Therefore, the features which are used to realize the registration of multi-source remote sensing image must keep invariant to image scale and rotation and be shown to provide robust matching across a substantial range of affine distortion and change in illumination.In this paper, scale space theory was studied, and several multi-scale feature detection algorithms which are often used currently were analyzed. After the comparison and appreciation to the several multi-scale feature detection algorithms, we draw the conclusion that SIFT algorithm proposed by LOWE (America,2004) is most suitable for the registration of multi-source remote sensing images. Then, advantages and disadvantages of SIFT algorithm were studied by experiments. Aiming the advantages and disadvantages, the current improved strategy of SIFT algorithm was analyzed. At last, the thought and process of local adaptive SIFT (LA-SIFT) automatic registration algorithm was built.Number of registration experiments were conducted by using LA-SIFT algorithm. The experimental results showed that the algorithm can realized highly precise and automatic registration of multi-source remote sensing images that possessed great gray-scale, perspective and resolution differences. At the same time, it made the matching feature points uniformly distribute in images, which further improved the accuracy of images geometric correction.Currently, automatic degree of remote sensing images registration is not high. Aiming the problem, LA-SIFT automatic registration system was designed. The system can realized the automatic and highly precise registration of multi-source remote sensing images, which ensured that remote sensing images registration could be realized speedily.