Research On Automatic Registration Of Optical And SAR Images

Optical remote sensing and SAR are currently two important technical means for obtaining surface information from ground observations.Due to the two kinds of data have complementary advantages,and they are widely used in the field of remote sensing image analysis.Image registration is required before using these two kinds of data comprehensively.Therefore,seeking a high-precision automatic registration method is a necessary prerequisite for achieving complementary advantages and comprehensive utilization of optical and SAR images.However,the imaging principles of optical remote sensing and SAR are completely different,resulting in a variety of differences between the acquired images,including imaging angle differences,ground resolution differences,and nonlinear intensity differences,this makes high-precision automatic registration between optical and SAR images difficult to register in heterogeneous images.Feature-based and area-based image registration methods are the current research hotspots in the field of image registration.In this paper,we investigate these two methods and propose a hybrid model to achieve high-precision automatic registration of large-size optical and SAR images by combining the advantages of both methods.The main research work in this paper is as follows.1.Firstly,the basic theory of image registration is analyzed and summarized from the three aspects of image registration principle,image registration method and image registration evaluation method.Then,the rotation invariance of the multi-source image matching algorithm RIFT is improved,which reduces the amount of calculation when constructing the rotationinvariant description vector.Secondly,based on the feature matching algorithm NBCS,a method which can automatically eliminate the matching with too large error and search the optimal homonym point in the neighborhood is added to improve the registration accuracy.2.The multi-modal reomote sensing image registration algorithm CFOG uses a differential gradient operator to express the geometric structure information of the image,but the differential gradient operator is easily affected by the speckle noise of the SAR image.In order to solve this problem,ROEWA operator which can effectively resist speckle noise is used instead of differential gradient operator to calculate the gradient of optical and SAR images.At the same time,in order to solve the problem of high overlap of template region when constructing dense feature expression,a method to reduce the overlap of template is proposed to reduce the repeated calculation.After multiple groups of image registration experiments,it is verified that the improved CFOG algorithm has better robustness and less calculation than CFOG algorithm.3.Aiming at the difficulty of achieving satisfactory registration results with a single registration method,combining the advantages of feature-based and region-based registration methods,a hybrid model-based registration method is proposed.Firstly,the improved RIFT algorithm is used to perform rough registration on the optical and SAR images,and then on the basis of this,the improved CFOG algorithm is used to further achieve fine registration.Finally,multiple sets of large-size optical and SAR images are used to perform registration experiments,and the verification is based on The hybrid model registration method can effectively overcome various differences between the optical and SAR images and achieve high-precision registration.