Research On SAR Image Registration Based On Improved SIFT And Deep Learning

After decades of long-term earth observation,remote sensing images have accumulated massive data,and some data are closely related to each other.It is of great significance for the study of large-scale spatio-temporal changes in many disciplines.Image splicing and image fusion of correlated remote sensing images are the necessary basis for obtaining relevant spatio-temporal data for various disciplines.Image registration is the basis of image splicing,image fusion,and the like,it is realized by placing the areas of geographic space overlap in the images to be spliced and to be merged into corresponding positions in the same coordinate axis.Due to its unique imaging mechanism,Synthetic Aperture Radar(SAR)has the features of all-time,all-weather,strong penetrability.Therefore,the related research on SAR images is of great scientific and practical importance.SIFT algorithm is a very classic algorithm in the image registration algorithm.At the same time,its various improved algorithms are also widely used in remote sensing images.Overcoming the influence of SAR image speckle noise on SIFT algorithm,improving SAR image registration accuracy and reducing the computing time are key issues in current SAR image registration,and is of important theoretical and practical significance.The feature detection modules,feature description modules and feature matching modules of the SIFT algorithm and the RANSAC algorithm are focused in this thesis.The specific study is as follows:1.On the basis of the SIFT-SSS algorithm,the first octave of Gaussian pyramids is not established in its feature detection.The first layer of Gaussian images of the second octave of pyramids can be obtained by simply taking the first S-layer Gaussian image points of the first pyramid every second pixel in each row and column.Therefore,the obtained Difference of Gaussian pyramid image reduces the effect of speckle noise,the number of detected key points is hardly affected,and the feature detection time is greatly reduced.2.The feature descriptor is generated using the SAE network without fine-tuning.Compared with the method of using PCA-SIFT to generate feature descriptor,the operation time for generating feature descriptor using the SAE network without fine-tuning to perform feature matching is significantly reduced.3.Using normalized cross-correlation as the similarity measure,and using the ratio of the second-largest/maximum-normalized cross-correlation coefficient to remove spurious interference point pairs during feature matching,compared with methods such as Euclidean distance measure and inverse cosine measure,and the like,the feature matching time is significantly reduced.4.For the problem that the probability of mis-matching points common in SAR image registration is much higher than 50%,considering the small variation of SAR image scale,the random sampling in each iteration is constrained to find a spatial transformation model.Therefore,the stability and accuracy of the registration result of RANSAC are improved,and the scale constraint reduces the probability of obtaining the correct model in each iteration,which greatly reduces the operation time.