Research On Optical And SAR Image Matching Method Based On Feature Learnin

The rapid development of Earth observation technology has made it possible to acquire remote sensing images of different modalities quickly.And fusing the different types of remote sensing image data to achieve complementary advantages has been widely used in many application scenarios such as change detection,military-target localization,etc.As a prerequisite and basis for these complex tasks,image matching aims to identify identical or similar structures and contents from two or more images,and to achieve spatial location consistency between images for the next step of joint processing and related applications.As a result,image matching has become a key research problem in many fields.Optical remote sensing and Synthetic Aperture Radar(SAR)are two important remote sensing techniques.The huge feature differences and nonlinear radiometric differences between their images greatly enhance the difficulty of matching,and traditional feature matching methods are less applicable to images with significant differences.With the rise and rapid development of deep learning technology,it has shown obvious superiority for image feature extraction and learning.Therefore,this paper focuses on the optical and SAR image matching problem,starting from the network learning deep features of images,and focuses on the optical and SAR image matching method based on feature learning,and the main research work is as follows:(1)Optical and SAR image matching network combined with SIFT and image patch feature learning.Combining the traditional SIFT method based on feature matching with deep learning methods to address the performance degradation of traditional feature matching methods in optical and SAR image matching.The underlying feature points with rotation and scale invariance are obtained by SIFT,and the corresponding image patches of optical and SAR images are intercepted.The image patch matching network is constructed to learn the image patch features and output the matched points,and finally the optical and SAR image matching results are optimised by combining the fast sample consistency method.(2)Optical and SAR image patch matching network combining multi-level keypoints detector and two-channel network structure.The traditional SIFT method has limited ability to obtain semantic information at high levels,making it incapable to obtain richer feature points from optical and SAR images.Therefore,a multi-level keypoints detector is proposed to obtain the low-level and high-level semantic information of the images by combining the weighted feature layers of the network output.And to filter the robust keypoints by constructing the peak value map in the local feature space and channel domain.Finally,optical and SAR image matching performance is improved by the two-channel image patch matching network.(3)Optical and SAR image matching method based on image translation.For feature differences that exist between optical and SAR images,the generative adversarial network is used for image translation.On the basis of circulation and symmetry,the pixel-level root mean square error loss is introduced,and using supervised learning to improve the quality of transformed images.The matching work is carried out between the generated optical image and the real optical image to reduce the difficulty of heterogenous image matching,which provides a new idea for optical and SAR matching.