A method and system for extracting same-name feature points of a space-borne SAR image

Abstract

The application discloses a kind of spaceborne SAR image homonymic feature point extraction method and system, the application is in view of SAR image high-precision matching demand, design a kind of end-to-end SAR image registration method based on convolutional neural network, through twin convolutional neural network architecture, realize the robust automatic extraction of spaceborne SAR image homonymic feature point.In addition, the method can also be popularly applied to optical image, hyperspectral image, night light image and the matching task of multiple types remote sensing image, is the key direction of current and even future image matching.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of microwave remote sensing, and relates to a spaceborne SAR image same-name feature point extraction method and system, in particular to a spaceborne SAR image same-name feature point extraction method based on a twin convolutional neural network. BACKGROUND

[0002] Spaceborne synthetic aperture radar (SAR) is an active imaging sensor for earth observation using microwaves, which has the advantages of all-weather, all-day, large range and strong penetration. However, due to the inherent side-looking slant-range imaging mechanism, the obtained image has geometric characteristics such as overlap, perspective shrinkage and shadow. When the SAR image pitch angle and azimuth angle change, the SAR image also shows great differences. In addition, the SAR imaging process is easily disturbed by noise, resulting in coherent speckle noise in the image. This makes it difficult to obtain reliable feature points in the SAR image matching process, and brings difficulties to the high-precision registration of SAR images.

[0003] In recent years, deep learning technology has developed rapidly. It is a completely data-driven solution that can automatically learn features from images. A lot of research has been done in the field of optical remote sensing image matching, but there is less research in the field of SAR image matching. Overall, there are two major challenges in SAR image matching: (1) lack of reliable SAR image registration data set; (2) lack of network model suitable for SAR image. SUMMARY

[0004] The present application proposes a spaceborne SAR image same-name feature point extraction method and system based on a twin convolutional neural network, which can effectively realize the robust extraction of SAR image same-name feature points, aiming at the problem that reliable same-name feature point extraction cannot be realized due to the influence of SAR image noise, geometric distortion and other factors in spaceborne SAR image matching.

[0005] The technical scheme adopted by the method of the present application is as follows: a spaceborne SAR image same-name feature point extraction method, comprising the following steps:

[0006] Step 1: obtaining a pair of spaceborne SAR images to be matched;

[0007] Step 2: inputting the pair of spaceborne SAR images to be matched into a spaceborne SAR image same-name feature point network to extract feature points of the spaceborne SAR images to be matched;

[0008] The spaceborne SAR image same-name feature point network is a twin network composed of two branches, each branch containing an input shared encoding module, a multi-scale convolution fusion module, a feature point detection decoding module and a feature descriptor detection decoding module.

[0009] The input shared encoding module comprises eight convolution layers with a convolution kernel size of 3 and a step size of 1 arranged in series, wherein the first four convolution layers have a size of 64, and the last four convolution layers have a size of 128; Group Normal normalization processing layers are added after the eight convolution layers; a maximum pooling layer with a size of 2*2 is added after the second layer, the fourth layer and the sixth layer of the convolution layers;

[0010] The multi-scale convolution fusion module is used for multi-scale feature fusion on the feature map F output by the input shared encoding module through four average pooling layers arranged in parallel; the first average pooling layer has a pooling kernel size of 1*1, the second average pooling layer has a pooling kernel size of 2*2, the third average pooling layer has a pooling kernel size of 3*3, and the fourth average pooling layer has a pooling kernel size of 4*4; a convolution layer with a convolution kernel of 1*1 and a deconvolution module are added after each average pooling layer to generate four feature maps z, x, c and v; the feature maps z, x, c and v are processed by Concat to generate multi-scale fusion features, and an overall feature map of final multi-scale fusion is generated through a ConBNReLU module;

[0011] The feature point detection decoding module is composed of two convolution layers, a Softmax layer and a Reshape layer arranged in series; the first convolution layer has a convolution kernel of 3*3 and a step size of 1, and the second convolution layer has a convolution kernel of 1*1 and a step size of 1; the Softmax layer performs softmax calculation on the feature point probability of the pixel by inputting the feature map of the two convolution layers; the Reshape layer reconstructs the feature single Tensor obtained by the Softmax layer, and outputs a one-dimensional feature point vector with the same size as the input image of the same feature point network of the spaceborne SAR image;

[0012] The feature descriptor detection decoding module is composed of two convolution layers, a Bi-CubicInterpolate layer and an L2 normalization layer arranged in series; the first convolution layer has a convolution kernel of 3*3 and a step size of 1, and the second convolution layer has a convolution kernel of 1*1 and a step size of 1; the Bi-CubicInterpolate layer inputs the 256-dimensional vector of the convolution layer to perform bilinear interpolation to expand the size; the L2 normalization layer performs L2 normalization processing on the 256-dimensional pixel descriptor to obtain a feature descriptor corresponding to the feature point vector output by the feature point detection decoding module.

[0013] The technical scheme adopted by the system of the application is as follows: a spaceborne SAR image same-name feature point extraction system comprising the following modules:

[0014] Module 1 is used for acquiring a spaceborne SAR image to be matched;

[0015] a module 2 for inputting a to-be-matched spaceborne SAR image pair into a spaceborne SAR image co-name feature point network to extract feature points of the to-be-matched spaceborne SAR image;

[0016] The spaceborne SAR image co-name feature point network is a twin network composed of two branches, and each branch includes an input shared encoding module, a multi-scale convolution fusion module, a feature point detection decoding module, and a feature descriptor detection decoding module.

[0017] The input shared encoding module includes eight convolution layers with a convolution kernel size of 3 and a step size of 1 connected in series, wherein the first four convolution layers have a size of 64, and the last four convolution layers have a size of 128; Group Normal normalization processing layers are added after the eight convolution layers; a maximum pooling layer with a size of 2x2 is added after the second layer, the fourth layer, and the sixth layer of the convolution layers;

[0018] The multi-scale convolution fusion module is used for multi-scale feature fusion of a feature map F output by the input shared encoding module through four average pooling layers connected in parallel; the first average pooling layer has a pooling kernel size of 1x1, the second average pooling layer has a pooling kernel size of 2x2, the third average pooling layer has a pooling kernel size of 3x3, and the fourth average pooling layer has a pooling kernel size of 4x4; a convolution layer with a convolution kernel of 1x1 and a deconvolution module are added after each average pooling layer to generate four feature maps z, x, c, and v; the feature maps z, x, c, and v are processed by Concat to generate multi-scale fusion features, and an overall feature map of final multi-scale fusion is generated through a ConBNReLU module;

[0019] The feature point detection decoding module is composed of two convolution layers, a Softmax layer, and a Reshape layer connected in series; the first 64-dimensional convolution layer of the two convolution layers has a convolution kernel of 3x3 and a step size of 1, and the second 65-dimensional convolution layer has a convolution kernel of 1x1 and a step size of 1; the Softmax layer performs softmax calculation on the feature point probability of the pixel by inputting the feature map of the two convolution layers; the Reshape layer reconstructs the feature single Tensor obtained by the Softmax layer to output a one-dimensional feature point vector with the same size as the input image of the spaceborne SAR image co-name feature point network;

[0020] The feature descriptor detection decoding module is composed of two convolutional layers, a Bi-CubicInterpolate layer and an L2 normalization layer; the first 256-dimensional convolutional layer of the two convolutional layers has a 3*3 convolution kernel and a step of 1, and the second 256-dimensional convolutional layer has a 1*1 convolution kernel and a step of 1; the Bi-CubicInterpolate layer inputs a 256-dimensional vector of the convolutional layer to perform bilinear interpolation to expand the size; and the L2 normalization layer performs L2 normalization processing on the 256-dimensional pixel descriptor to obtain a feature descriptor corresponding to the feature point vector output by the feature point detection decoding module.

[0021] The effective gain of the present application compared with the prior art is:

[0022] (1) In the star-borne SAR image same-name feature point network training data set preparation stage, the automatic sample making method of the SAR image registration data set self-supervision fusion multi-source feature is adopted, and the current SAR image registration data set based on deep learning is effectively solved. Compared with other deep learning of SAR image registration, the training book set of the method adopts multi-source feature information containing point features, geometric features and the like, realizes automatic extraction of the model to multiple features after training, and does not need additional multiple matching, so that the data processing efficiency is high.

[0023] (2) The star-borne SAR image same-name feature point network adopts a twin network architecture, realizes the flow processing of feature point detection, feature point description and feature matching, closely connects feature extraction and feature matching, and realizes the same-name feature point matching process considering image information.

[0024] (3) According to the imaging characteristics of the SAR image, in addition to the conventional feature detector loss and feature description generator loss, a descriptor difference maximization loss of SAR image noise and geometric distortion is designed, the loss realizes the elimination of the interference points extracted from the coherent speckle noise influence and geometric distortion area, guarantees the uniqueness of the feature descriptor, thereby reduces the influence of gross error points on feature matching, and realizes high-precision matching. BRIEF DESCRIPTION OF DRAWINGS

[0025] Figure 1 It is a star-borne SAR image same-name feature point network structure diagram of the embodiment of the present application;

[0026] Figure 2 It is a star-borne SAR image same-name feature point network training data set making flow chart of the embodiment of the present application.

[0027] Figure 3 It is a star-borne SAR image same-name feature point network training flow chart of the embodiment of the present application; DETAILED DESCRIPTION

[0028] For the convenience of those skilled in the art to understand and implement the present application, the present application will be further described in detail below in conjunction with the drawings and examples. It should be understood that the implementation examples described herein are only used to illustrate and explain the present application, and are not used to limit the present application.

[0029] The present application is a kind of spaceborne SAR image same name feature point extraction method, comprising the following steps:

[0030] Step 1: Obtain a pair of spaceborne SAR images to be matched;

[0031] Step 2: input the pair of spaceborne SAR images to be matched into the spaceborne SAR image same name feature point network, and extract the feature points of the spaceborne SAR images to be matched;

[0032] See Figure 1 The spaceborne SAR image same name feature point network of the present embodiment is a twin network composed of two branches, each branch contains an input shared encoding module, a multi-scale convolution fusion module, a feature point detection decoding module and a feature descriptor detection decoding module;

[0033] The input shared encoding module of the present embodiment includes eight convolution layers with a convolution kernel size of 3 and a step size of 1 connected in series, wherein the first four convolution layers have a size of 64 and the last four convolution layers have a size of 128; Group Normalization processing layers are added after the eight convolution layers of the present embodiment; 2x2 maximum pooling layers are added after the second, fourth and sixth convolution layers of the present embodiment;

[0034] The multi-scale convolution fusion module of the present embodiment is used for multi-scale feature fusion of the feature map F output by the input shared encoding module of the present embodiment through the parallel connection of four average pooling layers; the first average pooling layer of the present embodiment has a pooling kernel size of 1x1, the second average pooling layer has a pooling kernel size of 2x2, the third average pooling layer has a pooling kernel size of 3x3, and the fourth average pooling layer has a pooling kernel size of 4x4; a convolution layer with a convolution kernel of 1x1 and a deconvolution module are added after each average pooling layer of the present embodiment to generate four feature maps z, x, c and v; the feature maps z, x, c and v are processed by Concat to generate multi-scale fusion features, and a ConBNReLU module is used to generate the final multi-scale fusion feature map;

[0035] The feature point detection decoding module of the embodiment is composed of two convolutional layers, a Softmax layer and a Reshape layer; the first 64-dimensional convolutional layer of the two convolutional layers of the embodiment has a 3x3 convolution kernel and a step of 1, and the second 65-dimensional convolutional layer has a 1x1 convolution kernel and a step of 1; the Softmax layer of the embodiment performs softmax calculation on the feature point probability of the pixel by inputting the feature map of the two convolutional layers of the embodiment; the Reshape layer of the embodiment reconstructs the feature single Tensor obtained by the above Softmax layer, and outputs a one-dimensional feature point vector with the same size as the input image of the same-name feature point network of the spaceborne SAR image of the embodiment;

[0036] The feature descriptor detection decoding module of the embodiment is composed of two convolutional layers, a Bi-CubicInterpolate layer and an L2 normalization layer; the first 256-dimensional convolutional layer of the two convolutional layers of the embodiment has a 3x3 convolution kernel and a step of 1, and the second 256-dimensional convolutional layer has a 1x1 convolution kernel and a step of 1; the Bi-CubicInterpolate layer of the embodiment inputs the 256-dimensional vector of the convolutional layer of the embodiment to perform bilinear interpolation to expand the size; the L2 normalization layer of the embodiment performs L2 normalization processing on the 256-dimensional pixel descriptor to obtain a feature descriptor corresponding to the feature point vector output by the feature point detection decoding module of the embodiment.

[0037] The spaceborne SAR image same-name feature point network of the embodiment is a trained spaceborne SAR image same-name feature point network; the training process includes the following steps:

[0038] Step 2.1: Construct a training data set;

[0039] See Figure 2 , the specific implementation of step 2.1 includes the following sub-steps:

[0040] (1) Use the RPC geometric positioning model of the to-be-matched spaceborne SAR image pair data to solve the image coverage range, and solve the overlapping area range through the four corner point coordinates of the image coverage range of the to-be-matched spaceborne SAR image pair;

[0041] (2) According to the preset block size (2000x2000 is used in this method), the overlapping area of the to-be-matched spaceborne SAR image pair is processed without overlapping to obtain the initial to-be-matched image pair;

[0042] (3) The SIFT, SAR-SIFT and HOPC feature extraction methods are used in turn to extract the same-name feature points from the initial to-be-matched image pair;

[0043] (4) The same name feature points extracted by SIFT, SAR-SIFT and HOPC feature extraction methods in turn are gathered as two groups of one-to-one corresponding point sets;

[0044] (5) Based on the Euclidean distance as a similarity measure, a gross error elimination threshold is set, and the sample consistency algorithm is used to eliminate the gross error points. The gross error elimination threshold is set to 0.5, and finally the high-precision same name feature point pair (with a mean error of better than 0.5 pixels) is obtained;

[0045] (6) The high-precision same name feature point pairs of all initial image pairs to be matched are extracted in turn;

[0046] (7) All initial image pairs to be matched are cut without overlap according to the size of 256x256, and the image coordinates of the feature point pairs are converted into image coordinates in the 256 block coordinate system according to the block relationship of the images before and after cutting;

[0047] (8) Based on the 256 block coordinates, a binary label image is generated, and the feature point gray value is 1 and the other areas are 0;

[0048] (9) After generating all 256 size blocks of SAR image pairs in turn, the SAR image registration data set obtained by the SAR image pair, the same name feature point file and the label image is composed, that is, the SAR image matching data set for network training.

[0049] Step 2.2: Twin network training;

[0050] See Figure 3 , the pair of SAR images with the size of HxW and the corresponding label image are respectively taken as the input of the encoder E A and E B of the star-borne SAR image same name feature point twin network; through the encoder E A , the four levels of feature maps are finally spliced with the initial feature map to obtain a complete feature map; after obtaining the complete feature map, the feature maps of the SAR image pair are input into the corresponding decoder D A and D B to generate the final feature points and feature descriptors;

[0051] Three loss functions are used in the training process: feature point detector loss Loss p , descriptor loss Loss d , and descriptor difference maximization loss Loss differ ; when the error value of the loss function tends to be optimal, the training is completed.

[0052] The feature point detector loss Loss p of the embodiment is:

[0053]

[0054]

[0055] Wherein, X is the matching point feature map of the same-name feature point network identification of the spaceborne SAR image in this embodiment, and Y is the matching point feature map of the ground value of the corresponding same-name point pair in Label; H c W c These are the width and height of the feature map of the input image (256×256) of the spaceborne SAR image homonymous feature point network in this embodiment, with values ​​of [value missing]. h is the row number of a pixel in the feature map of this embodiment, w is the column number of a pixel in the feature map of this embodiment, and x hw y represents the feature point probability value of X at (h,w) in this embodiment. hw This represents the true probability value of the feature point Y at (h,w) in this embodiment; x hwy Represented as x hw The ground truth value in the y-th dimension, x hwn Represented as x hw The predicted value in the nth dimension; l p The purpose is to make the predicted probability value of the matching point identified by the same-name feature point network of the spaceborne SAR image in this embodiment as close as possible to the true value in Y.

[0056] The descriptor loss in this embodiment d for:

[0057]

[0058] l d (d w ,d' h’w’ ;if match )=λ d *if match *max(0,m p -d w T d' h’w’ ′)+(1-if match )*max(0,d w T d' h’w’ ′-m n )

[0059] Wherein, D is the matching descriptor feature map of the same-name feature point network identification of the spaceborne SAR image in this embodiment, and D′ is the matching feature descriptor feature map of the ground value of the corresponding same-name point pair in Label; if matchconsistency of homographic transformation of D and D' in this embodiment, taking value of 0 or 1, where 0 represents negative correlation and 1 represents positive correlation; d hw represents the descriptor feature value of D at (h, w), d' h′w′ represents the descriptor feature value of D' at (h', w'); λ d = 250 is a hyper parameter used to balance the positive correlation loss and negative correlation loss values within the feature point descriptor; m p = 1 is a positive correlation threshold value, m n = 0.2 is a negative correlation threshold value;

[0060] the descriptor difference maximization loss Loss differ of this embodiment is:

[0061]

[0062] wherein n represents the total amount of matching points and descriptors identified by the spaceborne SAR image homonymous feature point network of this embodiment for one image, i represents the i-th feature descriptor, j represents the j-th feature descriptor, and i≠j; represents the value of the i-th feature descriptor on the feature descriptor feature map, represents the value of the j-th feature descriptor on the feature descriptor feature map.

[0063] the loss function for extracting the spaceborne SAR image homonymous feature points of this embodiment is:

[0064] Total Loss = Loss p (X, Y) + Loss p (X', Y') + λLoss d (D, D') - αLoss differ (6)

[0065] wherein Loss p is the feature point detector loss; Loss d is the descriptor loss; Loss differ is the descriptor difference maximization loss; X is the matching point feature map identified by the spaceborne SAR image homonymous feature point network of this embodiment, Y is the matching point feature map of the true value of the corresponding homonymous point pair in Label; X' is the matching point feature map identified by the spaceborne SAR image homonymous feature point network of this embodiment, Y' is the matching point feature map of the true value of the corresponding homonymous point pair in Label; D is the matching descriptor feature map identified by the spaceborne SAR image homonymous feature point network of this embodiment, D' is the matching descriptor feature map of the true value of the corresponding homonymous point pair in Label; λ = 1 and α = 0.1 are hyper parameters for adjusting the proportion balance of the three loss functions.

[0066] In this embodiment, the specific implementation of step 2 includes the following steps:

[0067] (1) input the to-be-matched spaceborne SAR image into the trained spaceborne SAR image homonymic feature point network;

[0068] (2) the image pair generates corresponding homonymic feature points and feature descriptor vector files through the spaceborne SAR image homonymic feature point network;

[0069] (3) two groups of feature points and feature descriptors are obtained respectively, and feature point matching is performed based on the Euclidean distance as the similarity measure, and the corresponding threshold value can be set to eliminate gross error points;

[0070] (4) output the matched SAR image homonymic feature points, calculate the coordinates of the two groups of feature points, and calculate the transformation matrix of the SAR image pair;

[0071] (5) based on the calculated transformation matrix, resample the SAR image, and complete the SAR image matching task.

[0072] The present application aims at the high-precision matching requirement of SAR images, and designs an end-to-end SAR image registration method based on a convolutional neural network. Through a twin convolutional neural network architecture, robust automatic extraction of homonymic feature points of spaceborne SAR images is realized. In addition, this method can also be applied to the matching tasks of multiple types of remote sensing images such as optical images, hyperspectral images and night light images, and is the key direction of image matching at present and in the future.

[0073] In the data set construction stage, an automatic sample making method of SAR image registration data set self-supervision is realized. First, based on the RPC geometric positioning model, the overlapping area of the to-be-matched spaceborne SAR image pair data is divided, and an initial overlapping image pair is established. The traditional image feature extraction method is used for feature extraction on the obtained overlapping image pair. The present method uses SIFT, SAR-SIFT and HOPC feature extraction methods. When processing each overlapping image pair, all feature points extracted by different methods are collected, the Euclidean distance is used as the similarity measure, the sample consistency algorithm is used to remove gross error points, and a high-precision homonymic feature point pair is obtained. After the feature extraction of all overlapping image pairs is completed, the image pairs are cut without overlapping according to the size of 256x256, and the image coordinate conversion of the feature point pair is completed according to the block relationship of the image before and after cutting, and the corresponding binary Label image map is generated. Finally, the SAR image registration data set is composed of SAR image pairs, homonymic feature point files and Label image maps.

[0074] In the network training stage, a twin network architecture is adopted, and satellite-borne SAR image pairs A and B to be matched are processed respectively. Each branch network is composed of an encoder and a decoder. The encoder is responsible for generating a feature map containing global and local features of the SAR image. The decoder is divided into two parts, one is a feature detector for feature point detection, and the other is a feature description generator for feature point description. The loss of network training is composed of feature detector loss, feature description generator loss and descriptor difference maximization loss designed for SAR image characteristics.

[0075] In the test stage, the SAR image pair data to be matched is input, and the corresponding feature points of the SAR image pair are extracted based on the trained SAR image registration model. After the feature points extracted from the two SAR images, the Euclidean distance is used as the similarity measure for feature point matching. After removing the gross error points, reliable high-precision matching points are finally obtained.

[0076] It should be understood that the above description of the preferred embodiments is more detailed and should not be considered as limiting the scope of patent protection of the present application. Those skilled in the art can make substitutions or modifications without departing from the scope of protection of the present application, and all fall within the scope of protection of the present application. The scope of protection of the present application shall be subject to the appended claims.

Claims

1. A method for extracting homonymous feature points from spaceborne SAR images, characterized in that, Includes the following steps: Step 1: Obtain the spaceborne SAR image pairs to be matched; Step 2: Input the pair of spaceborne SAR images to be matched into the feature point network of the spaceborne SAR images, and extract the feature points of the spaceborne SAR images to be matched. The satellite-borne SAR image homonymous feature point network is a twin network consisting of two branches. Each branch includes an input shared coding module, a multi-scale convolutional fusion module, a feature point detection and decoding module, and a feature descriptor detection and decoding module. The input shared encoding module includes eight convolutional layers arranged in series, each with a kernel size of 3 and a stride of 1. The first four convolutional layers have a kernel size of 64, and the last four convolutional layers have a kernel size of 128. After each of the eight convolutional layers, a GroupNormal normalization layer is added. After the second, fourth, and sixth convolutional layers, a 2×2 max pooling layer is added. The multi-scale convolutional fusion module is used to perform multi-scale feature fusion on the feature map F output by the input shared encoding module through four parallel average pooling layers. The first average pooling layer has a 1×1 kernel size, the second average pooling layer has a 2×2 kernel size, the third average pooling layer has a 3×3 kernel size, and the fourth average pooling layer has a 4×4 kernel size. After each average pooling layer, a convolutional layer with a 1×1 kernel and a deconvolution module are added to generate four feature maps z, x, c, and v. The feature maps z, x, c, and v are processed by Concat to generate multi-scale fused features, and a ConBNReLU module is used to generate the final multi-scale fused overall feature map. The feature point detection and decoding module consists of two convolutional layers, a softmax layer, and a reshape layer cascaded together. The first 64-dimensional convolutional layer has a 3×3 kernel and a stride of 1, while the second 65-dimensional convolutional layer has a 1×1 kernel and a stride of 1. The softmax layer calculates the feature point probability of a pixel by using the feature maps from the two convolutional layers as input. The reshape layer reconstructs the feature single tensor obtained by the softmax layer and outputs a one-dimensional feature point vector with the same size as the input image of the satellite SAR image's corresponding feature point network. The feature descriptor detection and decoding module consists of two convolutional layers, a Bi-Cubic Interpolate layer, and an L2 normalization layer cascaded together. The first 256-dimensional convolutional layer has a 3×3 kernel and a stride of 1, while the second 256-dimensional convolutional layer has a 1×1 kernel and a stride of 1. The Bi-Cubic Interpolate layer takes the 256-dimensional vector from the convolutional layer as input and performs bilinear interpolation to enlarge its size. The L2 normalization layer performs L2 normalization on the 256-dimensional pixel descriptor to obtain a feature descriptor that corresponds one-to-one with the feature point vector output by the feature point detection and decoding module.

2. The method for extracting corresponding feature points from spaceborne SAR images according to claim 1, characterized in that: The satellite SAR image homonym network mentioned in step 2 is a trained satellite SAR image homonym network. Its training process includes the following steps: Step 2.1: Construct the training dataset; Step 2.2: Twin network training; A pair of SAR images of size H×W and their corresponding Label images are used as encoders for a twin network of corresponding feature points of spaceborne SAR images. and Input; via encoder Finally, the feature maps at the four levels are stitched together with the initial feature map to obtain a complete feature map. After obtaining the complete feature map, the feature maps of the SAR image pairs are input into their respective decoders. and The final feature points and feature descriptors are generated in the process; Three loss functions are used during training: feature point detector loss. Descriptor loss Descriptor difference maximization loss Training is complete when the error value of the loss function approaches the optimum.

3. The method for extracting corresponding feature points from spaceborne SAR images according to claim 2, characterized in that, Step 2.1 includes the following sub-steps: (1) The coverage area of ​​the image is calculated by using the RPC geometric positioning model of the satellite SAR image to be matched, and the overlapping area is calculated by using the coordinates of the four corner points of the coverage area of ​​the satellite SAR image to be matched. (2) Perform non-overlapping block processing on the overlapping areas of the spaceborne SAR image pairs to be matched according to the preset block size to obtain the initial image pairs to be matched. (3) For the initial image pairs to be matched, SIFT, SAR-SIFT and HOPC feature extraction methods are used in sequence to extract corresponding feature points for the initial image pairs to be matched. (4) Collect the corresponding feature points extracted by SIFT, SAR-SIFT and HOPC feature extraction methods in sequence to form two sets of points that correspond one-to-one; (5) Based on Euclidean distance as a similarity measure, a threshold for removing outliers is set, and outliers are removed using the sample consistency algorithm to finally obtain high-precision pairs of identical feature points. (6) Sequentially extract high-precision corresponding feature point pairs for all initial image pairs to be matched; (7) All initial image pairs to be matched are cropped without overlap according to the size of N×N. Based on the block relationship of the images before and after cropping, the image coordinates of the feature point pairs are converted into image coordinates in the N block coordinate system; where N is a preset value. (8) Based on N block coordinates, generate a binary Label image map, with feature point grayscale value of 1 and other areas of 0; (9) After generating all N-sized SAR image pairs in sequence, the obtained SAR image registration dataset consists of SAR image pairs, corresponding feature point files and Label image maps, which is the SAR image matching dataset trained by the network.

4. The method for extracting corresponding feature points from spaceborne SAR images according to claim 2, characterized in that: In step 2.2, the feature point detector loses... for: （1） （2） in, The matching point feature map identified by the same-name feature point network of the spaceborne SAR image. The feature map of matching points for the ground values ​​of the corresponding points in the Label; These are the width and height of the feature map of the input image of the satellite-borne SAR image homonymous feature point network, with values ​​of [value missing]. ; It is the row number of a certain pixel in the feature map of the feature points. It is the column number of a certain pixel in the feature map of the feature points. Indicates the exist The probability value of the feature point at that location. Indicates the exist True probability values ​​of feature points at a given location; Represented as In the The ground truth value in each dimension, Represented as In the Predicted values ​​for each dimension; The purpose is to make the predicted probability values ​​of matching points identified by the homonymous feature point network of the spaceborne SAR image as close as possible. The truth value in; The descriptor loss for: （3） in, The matching descriptor feature map identified by the network of corresponding feature points in the spaceborne SAR image. The feature descriptor feature map for matching the ground values ​​of the corresponding points in the Label; Indicates the and The homography transformation consistency is represented by a value of 0 or 1, where 0 indicates negative correlation and 1 indicates positive correlation. express exist Descriptor feature values ​​at the location, express exist Descriptor feature values ​​at the location; It is a hyperparameter used to balance the positive and negative correlation loss values ​​within the feature point descriptor; The positive correlation threshold The threshold for negative correlation; The descriptor difference maximization loss for: （5） in, This represents the total number of matching points and descriptors identified by the satellite-borne SAR image homonym network for a given image scene. Indicates the first A feature descriptor, Indicates the first Each feature descriptor, and ; Indicates the first The values ​​of each feature descriptor on the feature descriptor feature map Indicates the first The values ​​of each feature descriptor on the feature descriptor feature map; The loss function for extracting the corresponding feature points from the spaceborne SAR image is: (6) in, For feature point detector loss; For descriptor loss; Maximize the loss for descriptor differences; The matching point feature map identified by the same-name feature point network of the spaceborne SAR image. The feature map of matching points for the ground values ​​of the corresponding points in the Label; The matching point feature map identified by the same-name feature point network of the spaceborne SAR image. The feature map of matching points for the ground values ​​of the corresponding points in the Label; The matching descriptor feature map identified by the network of corresponding feature points in the spaceborne SAR image. The feature descriptor feature map for matching the ground values ​​of the corresponding points in the Label; , As a hyperparameter, the weighting of the three loss functions is adjusted to balance them.

5. The method for extracting corresponding feature points from spaceborne SAR images according to any one of claims 1-4, characterized in that: Step 2 includes the following steps: (1) Input the spaceborne SAR image to be matched into the trained spaceborne SAR image homonym network; (2) Image pairs generate corresponding feature points and feature descriptor vector files through the satellite-borne SAR image homonymous feature point network; (3) Using the two sets of feature points and feature descriptors obtained respectively, feature points are matched based on Euclidean distance as a similarity measure, and gross errors can be removed by setting the corresponding threshold. (4) Output the matching SAR image with the same name feature points.

6. The method for extracting corresponding feature points from spaceborne SAR images according to claim 5, characterized in that: Based on the corresponding feature points of the matched SAR images, the coordinates of the two sets of feature points are calculated to calculate the transformation matrix of the SAR image pair; based on the calculated transformation matrix, the SAR images are resampled to complete the SAR image matching task.

7. A system for extracting homonymous feature points from spaceborne SAR images, characterized in that, Includes the following modules: Module 1 is used to acquire the spaceborne SAR image to be matched; Module 2 is used to input the pair of spaceborne SAR images to be matched into the feature point network of the spaceborne SAR images and extract the feature points of the spaceborne SAR images to be matched. The satellite-borne SAR image homonymous feature point network is a twin network consisting of two branches. Each branch includes an input shared coding module, a multi-scale convolutional fusion module, a feature point detection and decoding module, and a feature descriptor detection and decoding module. The input shared encoding module includes eight convolutional layers arranged in series, each with a kernel size of 3 and a stride of 1. The first four convolutional layers have a kernel size of 64, and the last four convolutional layers have a kernel size of 128. After each of the eight convolutional layers, a GroupNormal normalization layer is added. After the second, fourth, and sixth convolutional layers, a 2×2 max pooling layer is added. The multi-scale convolutional fusion module is used to perform multi-scale feature fusion on the feature map F output by the input shared encoding module through four parallel average pooling layers. The first average pooling layer has a 1×1 kernel size, the second average pooling layer has a 2×2 kernel size, the third average pooling layer has a 3×3 kernel size, and the fourth average pooling layer has a 4×4 kernel size. After each average pooling layer, a convolutional layer with a 1×1 kernel and a deconvolution module are added to generate four feature maps z, x, c, and v. The feature maps z, x, c, and v are processed by Concat to generate multi-scale fused features, and a ConBNReLU module is used to generate the final multi-scale fused overall feature map. The feature point detection and decoding module consists of two convolutional layers, a softmax layer, and a reshape layer cascaded together. The first 64-dimensional convolutional layer has a 3×3 kernel and a stride of 1, while the second 65-dimensional convolutional layer has a 1×1 kernel and a stride of 1. The softmax layer calculates the feature point probability of a pixel by using the feature maps from the two convolutional layers as input. The reshape layer reconstructs the feature single tensor obtained by the softmax layer and outputs a one-dimensional feature point vector with the same size as the input image of the satellite SAR image's corresponding feature point network. The feature descriptor detection and decoding module consists of two convolutional layers, a Bi-Cubic Interpolate layer, and an L2 normalization layer cascaded together. The first 256-dimensional convolutional layer has a 3×3 kernel and a stride of 1, while the second 256-dimensional convolutional layer has a 1×1 kernel and a stride of 1. The Bi-Cubic Interpolate layer takes the 256-dimensional vector from the convolutional layer as input and performs bilinear interpolation to enlarge its size. The L2 normalization layer performs L2 normalization on the 256-dimensional pixel descriptor to obtain a feature descriptor that corresponds one-to-one with the feature point vector output by the feature point detection and decoding module.

Images