Search

CN-122023469-A - Parallax interference resistant structured edge guide registration method, system, equipment and storage medium for high maneuvering satellite remote sensing image

CN122023469ACN 122023469 ACN122023469 ACN 122023469ACN-122023469-A

Abstract

A method, a system, equipment and a storage medium for structured edge guide registration of a high maneuvering satellite remote sensing image with parallax interference resistance belong to the technical field of remote sensing image processing. The method comprises the steps of obtaining a high maneuvering satellite double-phase remote sensing image to be registered, generating a structure prior probability map, guiding a characteristic extraction network to focus on ground feature structural characteristics by using the probability map, respectively predicting an affine transformation matrix representing global rigid transformation and a projection transformation matrix representing local nonlinear deformation by using a hierarchical strategy, mapping the transformation matrix into a displacement vector field by using a space transformation network, extracting pixel-level residual displacement by calculating the difference value between the projection displacement field and a standard basic grid, superposing to obtain a final displacement vector field, and sampling and registering a source image according to the final displacement vector field. By fusing structure edge knowledge and displacement field residual error mechanisms, the problems of nonlinear distortion and large visual angle parallax interference in high maneuver imaging are effectively solved, and high-precision and high-robustness registration in complex scenes is realized.

Inventors

  • XUE MINGLIANG
  • FENG LIN

Assignees

  • 大连民族大学

Dates

Publication Date
20260512
Application Date
20260123

Claims (9)

  1. 1. The parallax interference resistant structured edge guide registration method for the high-mobility satellite remote sensing image is characterized by comprising the following steps of: S1, acquiring a source image to be registered And a target image The image is a double-time-phase remote sensing image acquired by a high mobility satellite; S2, extracting the feature based on edge knowledge guidance, namely extracting explicit structure edge information of the source image and the target image by using an edge detection operator to generate a structure edge image And Mapping the structure edge image into a structure prior probability map, and carrying out weighted guidance on the original image characteristics by utilizing the probability map to obtain fusion characteristics focused on the ground object structure And ; S3, hierarchical transformation matrix prediction, namely processing the fusion characteristics by utilizing an affine matrix prediction module to predict an affine transformation matrix representing global rigid transformation; s4, a residual error fusion step of a displacement vector field, wherein the affine transformation matrix is mapped into an affine displacement vector field by utilizing a space transformation network Mapping the projective transformation matrix into a projective displacement vector field Calculating the difference between the projected displacement vector field and the standard basic grid to extract pixel level residual displacement, and superposing the residual displacement with the affine displacement vector field to obtain a final displacement vector field ; S5, registering based on the final displacement vector field And sampling and interpolating the source image to obtain a registered image.
  2. 2. The method of claim 1, wherein the edge knowledge-based feature extraction step comprises computing image gradient magnitudes using a Sobel operator to obtain a structural edge image And Inputting the structural edge image into an edge attention module, extracting deep edge features through a backbone network, generating a structural prior probability map representing the confidence that pixel points belong to the structural edge through an activation function, performing point multiplication fusion on the structural prior probability map and the original image features extracted by the feature extraction backbone network, and generating fusion features with structural attention weights And 。
  3. 3. The method according to claim 1, wherein the hierarchical transformation matrix prediction step comprises the affine transformation matrix prediction module receiving fusion features And Affine transformation parameters of the freedom degree of regression prediction through a convolution layer and a full connection layer, wherein the projection transformation matrix prediction module receives the target image characteristics And the source image characteristics transformed by affine transformation parameters are processed by projective transformation parameters of the convolution layer and the full-connection layer regression prediction freedom degree so as to process oblique parallax and perspective deformation.
  4. 4. The method according to claim 1, wherein in the displacement vector field residual fusion step, the final displacement vector field is The calculation formula of (2) is as follows: ; Wherein, the For an affine displacement vector field, For the projection of the displacement vector field, A standardized coordinate grid of the same size as the displacement field; The pixel level residual displacement is characterized as purely nonlinear.
  5. 5. The method of claim 1, further comprising a model training step based on a composite supervised loss function, the loss function Comprising the following steps: ; Wherein, the The geometric distance loss of the key points is used for restraining the alignment precision of the sparse key points selected based on the edge knowledge; Is a bi-directional registration consistency loss, used for constraining reversibility of forward transformation and reverse transformation; for structuring mutual information loss, the intensive consistency of the registration image and the target image on the semantic structure is constrained by minimizing the joint entropy of the feature space, so that supervision training under weak annotation is realized.
  6. 6. The method of claim 2, wherein the edge attention module employs ResNet as a backbone network, the feature extraction backbone network employs ResNet, and the activation function is a Sigmoid or SoftMax function.
  7. 7. A parallax interference resistant structured edge-guided registration system for high mobility satellite remote sensing images, comprising: The acquisition module is used for acquiring a high mobility satellite double-phase remote sensing image to be registered; The edge guide feature extraction module is used for extracting explicit structural edge information of the image and generating a structural prior probability map, and the probability map is used for guiding the feature extraction network to focus on the structural features so as to obtain fusion features; The hierarchical matrix prediction module is used for predicting an affine transformation matrix representing global transformation and a projective transformation matrix representing local nonlinear deformation respectively; The displacement field residual error fusion module is used for carrying out difference between the displacement field generated by the projection transformation matrix and the basic grid to obtain residual error displacement, and fusing the residual error displacement with the displacement field generated by the affine transformation matrix to generate a final displacement vector field; and the registration module is used for carrying out registration processing on the source image based on the final displacement vector field.
  8. 8. A computer device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the method of any of claims 1 to 6 when the computer program is executed.
  9. 9. A computer readable storage medium storing computer instructions which, when invoked, are adapted to perform the method of any one of claims 1 to 6.

Description

Parallax interference resistant structured edge guide registration method, system, equipment and storage medium for high maneuvering satellite remote sensing image Technical Field The invention relates to the technical field of remote sensing image processing, in particular to a parallax interference resistant high maneuvering satellite remote sensing image structured edge guiding registration method, a system, equipment and a storage medium. Background The remote sensing image registration is a key technology for performing geometric alignment on remote sensing images acquired by different time, different sensors or different visual angles under a unified coordinate system, and is a precondition of subsequent image fusion and change detection. With the development of remote sensing technology, high mobility satellites are widely used. However, such dual phase image registration acquired by satellites faces the great challenge that nonlinear geometric distortion, rapid changes in satellite pose, introduce complex geometric deformations, including not only rigid transformations, but also severe local nonlinear stretching. The local radiation distortion caused by cloud and fog shielding, illumination change and the like causes the failure of the traditional registration method based on gray scale. And the high-rise building presents distinct projection forms under different visual angles, so that the features of the ground objects are not overlapped, and mismatching is easy to generate. In the prior art, the traditional registration method often depends on global transformation, local nonlinear deformation is ignored, and the common deep learning method lacks explicit utilization of physical structural features and is easy to fail in texture missing or repeated areas. Therefore, a method is needed that can effectively use the structural edge knowledge and can handle complex nonlinear deformations through displacement field fusion. Disclosure of Invention The invention aims to provide a parallax interference resistant structured edge guide registration method for a high maneuvering satellite remote sensing image, which is used for realizing high-precision registration under a complex scene by processing nonlinear deformation through fusion affine and projection displacement vector fields and guiding key point selection by combining structural edge knowledge. The technical scheme adopted by the invention is that the high-mobility satellite remote sensing image structured edge guide registration method for resisting parallax interference comprises the following steps: S1, acquiring a source image to be registered And a target imageThe image is a double-time-phase remote sensing image acquired by a high mobility satellite; S2, extracting the feature based on edge knowledge guidance, namely extracting explicit structure edge information of the source image and the target image by using an edge detection operator to generate a structure edge image AndMapping the structure edge image into a structure prior probability map, and carrying out weighted guidance on the original image characteristics by utilizing the probability map to obtain fusion characteristics focused on the ground object structureAnd; S3, hierarchical transformation matrix prediction, namely processing the fusion characteristics by utilizing an affine matrix prediction module to predict an affine transformation matrix representing global rigid transformation; s4, a residual error fusion step of a displacement vector field, wherein the affine transformation matrix is mapped into an affine displacement vector field by utilizing a space transformation network Mapping the projective transformation matrix into a projective displacement vector fieldCalculating the difference between the projected displacement vector field and the standard basic grid to extract pixel level residual displacement, and superposing the residual displacement with the affine displacement vector field to obtain a final displacement vector field; S5, registering based on the final displacement vector fieldAnd sampling and interpolating the source image to obtain a registered image. Further, the feature extraction step based on edge knowledge guidance specifically includes calculating an image gradient amplitude by using a Sobel operator to obtain a structural edge imageAndInputting the structural edge image into an edge attention module, extracting deep edge features through a backbone network, generating a structural prior probability map representing the confidence that pixel points belong to the structural edge through an activation function, performing point multiplication fusion on the structural prior probability map and the original image features extracted by the feature extraction backbone network, and generating fusion features with structural attention weightsAnd。 Further, the hierarchical transformation matrix prediction step specifically includes the affine transforma