| Compared with optical and infrared sensors,Synthetic Aperture Radar(SAR)is not limited by light and weather conditions,and can achieve active observation of the Earth around the clock.SAR image registration is one of the basic technologies of image processing.The goal of registration is to find the corresponding relationship between two or more SAR images,and align them in the same coordinate system,so that they have the same geometric space information,so as to achieve the comparison and analysis of SAR images at different times or under different conditions,so as to achieve subsequent applications,such as change detection,image fusion and target detection.Therefore,the study of SAR image registration has a strong practical significance.The repeatability of keypoints and the effectiveness of feature descriptors directly affect the accuracy of feature based registration methods.This paper starts from two aspects of SAR image keypoint detection and feature descriptor extraction,and at the same time considers the impact of SAR image speckle noise,carries out research on SAR image registration methods for complex scenes based on feature intersection,fully excavates the local features of highly repetitive keypoints and establishes reliable matching point pairs,Thus achieving accurate registration of SAR images.The main work of this article is as follows:In response to the difficulty of feature extraction in SAR images,the low number and stability of extracted effective features,and the inability of a single feature to meet the requirements of extracting highly repetitive keypoints in complex scene SAR images.In this paper,a novel keypoint detector based on feature intersection is proposed,which includes three parallel detectors: using phase consistency to obtain phase consistency feature maps on logarithmic SAR images,and using the minimum moment method to obtain image corners;Using structural tensors to detect candidate keypoints in SAR images with significant horizontal and vertical gradient changes;Finally,a local variation coefficient keypoint detector is used to detect candidate keypoints outside the large geometric distortion and strong scattering regions,thereby avoiding erroneous matching in complex regions.This algorithm can effectively extract keypoints with high repeatability,and greatly reduce the number of erroneous keypoints,thereby reducing the computational cost of feature description and matching.Aiming at the problem of difficulty in extracting feature descriptors from keypoints of SAR images and insufficient discrimination of feature descriptors.This paper designs a siamese cross stage partial network for extracting deep feature descriptors from keypoints.This network enhances feature reuse through dense connections,fully utilizes feature information from different convolutional layers,and uses cross stage partial connections between dense connection blocks and transition layers,reducing the amount of computation that increases as the network deepens,and improving the speed of generating feature descriptors,Compared to traditional hand designed shallow descriptors,it can be used to obtain more accurate matching point pairs. |
