Research On Point Pattern Chamfer Registration Of Optical Image And SAR Image Based On Mesh Grids

Image registration is the process of matching and overlaying two similar images which captured by different sensors or acquired under different external conditions.Compared with single-sensor image,the multiple-sensor image registration can provide researchers with more comprehensive information.Synthetic aperture radar(SAR)is an active microwave imaging system,which can acquire images all-weather and allday.Compared with the optical sensor,SAR has the ability of penetrating some ground objects,while optical image has richer spectral information than SAR image.Optical image and SAR image registration can make use of the advantages of both and obtain more abundant information.Therefore,optical-to-SAR image registration is of great significance and has been applied in many fields,such as computer vision and remote sensing image processing.However,the differences of geometric and radiotric characteristics between optical and SAR image,and the large amount of speckle noises in SAR image make the optical-to-SAR image registration difficult.The existing optical-to-SAR image registration methods can be divided into two categories: the intensity-based method and the feature-based method.However,each method has its own shortcomings.The intensity-based registration methods need a lot of computation and can not solve the deformation between the images.The featurebased registration methods have difficulties in finding the corresponding features between optical image and SAR image.Therefore,there is no algorithm to solve all the problems of optical image and SAR image registration.This paper proposes a point pattern chamfer registration method for optical image and SAR image based on mesh grids,which constructs a coarse to fine matching framework.There are three main contributions in this paper:(1)A modified holistically edge detection algorithm based on convolutional neural network is proposed to extract main contour edge features of optical image and SAR image,which overcomes the interference of speckle noises and detailed edges.Therefore,the algorithm proposed in this paper can achieve the registration of images with long acquisition time interal and large resolution differences.(2)Mesh grid strategy is proposed to introduce point pattern chamfer mathing method,build a coares to fine matching framework,make up for the shortcomings of fast directional chamfer matching and realize the selection the intergration of maching point pairs.The mesh grid strategy can complete the maching of main contour edges and solve the problem of slight contour deformation,so as to lay a foundation for the subsequent non-linear image alignment.(3)Moving direct linear transform is introduced to reduce the parallax between the optical image and SAR image instead of a homography matrix.The experiments are carried out on five groups of real optical image and SAR image,the algorithm proposed in this paper can complete the registration of the five sets of data.The experimental results show the accuracy and robustness of our proposed method.