Research On Registration Method For Remote Sensing Images

The registration of remote sensing images is the first-phase preparations for further processing of mosaic,fusion,and stitching.Therefore,improving the registration accuracy of remote sensing images is a key research area in this field.High-resolution satellite images have high resolution and rich information,covering a large range of ground and reflecting the characteristics of ground information in detail.Registration of high-resolution remote sensing images imposes higher requirements on remote sensing image registration methods.Single image registration method can no longer meet the requirements of remote sensing image registration.In image registration,registration methods based on straight line features have disadvantages such as inaccurate positioning information and undetected straight lines.This thesis proposes a remote sensing image registration method based on the combination of straight line features and regional features.This method uses Hough transform to detect linear features in remote sensing images,remote sensing image main direction to calculate rotation parameters,optimal threshold segmentation method to segment remote sensing images and Hu moment invariant to be the measure of the similarity measure of the sub-region block of remote sensing image which not only plays an important role in the accuracy of straight line orientation in image registration,but also solves the problem of inaccurate positioning of the straight line.The new method can be applied to the translation,rotation and scaling of remote sensing images improving the registration accuracy of remote sensing imagesIn the traditional SIFT-RANSAC based image registration methods,firstly,the SIFT algorithm is used to perform rough registration on the image,and then the RANSA algorithm is used to eliminate the feature points that are not meaningful.In practical applications,since RANSAC has limitations in removing non-meaningful feature points,it may mistakenly remove key feature points.This thesis proposes an SIFT-RANSAC bases remote sensing image registration optimization method.Firstly,the histogram statistics of the extracted feature points are made.Then,according to the statistical results of the histogram,the remote sensing image is compensated for angles,eliminating some feature points having mistakes in some angles.The feature points provide a more accurate sample database for the next RANSAC parameter estimation model.This method alleviates the situation of mistaking to eliminate characteristic points to some extent,thus improving the accuracy of remote sensing image registration.In this thesis,using the captured images of high-resolution satellites taken in Heilongjiang Province in 2017 as experimental data.several groups of feature points extraction and image registration experiments were performed.Correct feature points and incorrect feature points were extracted and the correct matching rate during the registration process was calculated.The rate was compared with the unoptimized method and the experiment verified the validity and accuracy of the new proposed registration method.