Registration Technique For High-resolution Remote Sensing Images Based On Feature Points

In the area of remote sensing application, image registration is an important step for image fusion. change detection, image correction and image mosaic. Remote Image Registration is defined that two or more remote images from different time or different sensors or different perspective of the same scenery. Even though there are many existing Image Registration Technology, some of which have obtained business applications. but they are not able to cope with remote images owing to rich information and much noise. Therefore, it is most necessary to studying on Remote Image Registration Automatic Technology.Among Image Registration algorithm, the algorithm based on SIFT has been one of widely applied Image Registration algorithm thanks to this fact that SIFT features are characteristic of fine distinctiveness,rich information, keeping invariant to image scaling and rotation, and partially invariant to change in illumination. In order to effectively eliminate the mismatched point pair, this algorithm based on SIFT often combined with Random Sample Consensus (RANSAC) algorithm to realize called "coarse+fine" matching model. However, it is still unable to meet the requirements of remote sensing image registration. The experimental results show that that it takes too much time in running the program, and sensitive to image noise points, its matching effect is bad.For high-resolution remote images with tremendous feature, a new method of image registration is proposed by combining contourlet transform with SIFT. Firstly, the discrete contourlet transform decomposes the original image into a low-frequency sub-band and several high-frequency sub-bands. For high-frequency sub-bands, image edge feature points are extracted by an appropriate threshold. For the low-frequency sub-band, image SIFT feature are extracted by SIFT. Then, coarse control point pairs are searched out with normalized cross-correlation matching and SIFT matching before accurate control point pairs are hunted by RANSAC (Random Sample Consensus). The experimental results show that, as for high-resolution remote sensing image registration, this method can extract more accurately control point pairs,own higher computing efficiency and matching ratio than these methods based on NSCT or SIFT.In addition, I have improved the acquisition strategy of feature points and the selection strategy of the best matching control point pair set. To a certain extent, so I make RANSAC algorithm better. The experimental results show that, improved RANSAC performs better than traditional RANSAC on the running time and the number of matching point pair.