Cross-viewpoint geographic positioning method and system based on dense partitioning strategy

Abstract

The present invention discloses a cross-viewpoint geolocation method and system based on a dense partitioning strategy. The geolocation network model constructed by the present invention includes a drone branch and a satellite image branch. The two branches adopt a shared weight learning method, and each branch includes a feature extraction part, a feature partitioning processing part, and a classification supervision part. The initial input of the network is a given image X, and the corresponding branch is selected according to the image perspective type; the given image X is subjected to feature extraction through a backbone network to obtain high-dimensional features; the high-dimensional feature map is partitioned using a dense partitioning strategy to obtain partition blocks; then the partition blocks are averaged and pooled to obtain compressed features; a classifier module is used to predict a geographic tag for each compressed feature, and the cross entropy loss is used to calculate the difference, thereby optimizing the network model. The present invention enhances the robustness of the algorithm to the size and position offset of objects in the image from both local and global levels, and improves the accuracy of multi-viewpoint image matching.

Description

Technical Field

[0001] The present invention relates to the field of image retrieval technology, and in particular to a cross-viewing angle geographic positioning method and system based on a dense partitioning strategy. Technical Background

[0002] Cross-view geolocation is a challenging computer vision problem. Since images from different perspectives may have large appearance changes, such as lighting, occlusion, geometric deformation, etc., these factors will directly interfere with the performance of the cross-view geolocation algorithm. In order to solve this problem, some methods based on deep learning have been proposed, which can be mainly divided into two categories: feature point-based methods and feature region partitioning-based methods. Among them, the feature region partitioning-based method refers to dividing the image features into several region blocks, and then matching them based on the similarity between the region blocks. This type of method has low computational complexity and good context information expression ability, but it also has some disadvantages. For example, the size, number of partitions, and shape of the divided feature region blocks will affect the performance of the cross-view geolocation algorithm. Secondly, the position distribution and overall size proportion of the retrieval target in the image are not fixed. The partitioning strategy with a fixed center position has the problem of dividing the same continuous feature into multiple feature region blocks. Summary of the invention

[0003] In view of the shortcomings of the prior art, the present invention provides a cross-view geolocation method and system based on a dense partitioning strategy, which is used for cross-view geolocation tasks by designing a dense partitioning strategy. The dense partitioning strategy obtains semantic information blocks that are independent of the image direction and are not affected by the size and spatial distribution of the retrieval target by performing multi-scale and multi-shape cutting and partitioning on image features. The dense partitioning strategy can extract contextual information between a wider range of regions, solve the problem of segmenting continuous features into different regions, and alleviate the problem of inaccurate retrieval caused by the spatial position offset and scale uncertainty of the retrieval target. The dense partitioning strategy proposed in this patent can significantly improve the accuracy of the cross-view geolocation algorithm.

[0004] The technical solution adopted by the present invention to solve the technical problem is as follows:

[0005] The cross-view geolocation method based on dense partitioning strategy includes the following steps:

[0006] Step 1: Dataset creation: Create two image datasets required for network training.

[0007] Download the public dataset University-1652 dataset. The original downloaded University-1652 dataset is a UAV-based geolocation dataset. It contains data from three platforms, namely, images of 1,652 university buildings from drone, satellite, and ground perspectives. The drone-view images contained in the dataset have less content occlusion, better visibility, and both belong to aerial perspectives with satellite views, have similar structures, and are easier to match with satellite images. This dataset is used for two tasks, namely, target positioning from the drone perspective and drone navigation. The training set includes 701 buildings from 33 universities, and the test set includes the other 951 buildings from the remaining 39 universities. There are no overlapping universities in the training set and the test set. The drone and satellite perspectives in this dataset are used for drone view target positioning. There are 37,855 drone view images in the query set, 701 true matching satellite view images and 250 satellite view interference items in the dataset, and for each drone view image, there is only one true matching satellite view image. The goal of the drone geolocation task is to find the corresponding satellite view image based on the drone view image, so as to determine the geographic location of the drone.

[0008] Step 2: Construct a cross-view geolocation network model based on dense partitioning strategy;

[0009] The geolocation network model consists of a drone branch and a satellite image branch, and the two branches adopt a shared weight learning method. Each branch consists of three parts: feature extraction part, feature partitioning processing part and classification supervision part.

[0010] Specific structure of feature extraction part: This paper uses SwinV2 as the backbone network to extract visual features, and uses parameters trained based on the ImageNet-21K dataset as the pre-trained weights of the backbone network. Specifically, given an input image X with a size of 3×H×W, after the backbone network extracts features, a shape of The high-dimensional feature map f i , the corresponding operation is expressed as:

[0011] f i =F swinV2 (X i ),i∈{D,S} (1)

[0012] Wherein, the subscripts D and S represent the drone view and satellite view, respectively, H represents the height of the input image X, and W represents the width of the input image X; F swinV2 Represents the feature extraction operation of the backbone network.

[0013] Specific structure of feature partitioning processing: Design dense partitioning strategy for high-dimensional feature graph fi Divide and obtain N (in order to explain the structure of the patent in detail, the present invention defines N = 10) divided blocks with different shapes. First, the high-dimensional feature map f i Slice from the center to the outside into four non-overlapping annular areas, and define the annular areas as basic partition blocks P1, P2, P3 and P4 respectively. The specific steps are as follows:

[0014] (a) Calculate the size increment Δh of each basic partition block,

[0015]

[0016] Among them, floor() means that the result is rounded down.

[0017] (b) For the high-dimensional feature map f i After partitioning, four basic partition blocks P1, P2, P3 and P4 are obtained. The partitioning order is as follows:

[0018]

[0019] The four basic partition blocks P1, P2, P3 and P4 obtained do not have overlapping areas. On this basis, by densely combining these basic partition blocks, 10 partition blocks with different shapes are obtained, such as Figure 3 As shown. The shape of each partition block is expressed as follows:

[0020]

[0021] Among them, ■ represents a square partition (solid); except for the square partition, other partitions are square ring partitions, that is, the difference between the large square partition and the small square partition (hollow).

[0022] like Figure 3 As shown in the figure, the combined partition blocks can expand the boundaries of the basic partition blocks to varying degrees, thereby alleviating the problem of continuous features being divided into different regions. Then the average pooling layer is used to compress the spatial dimensions of the partition blocks so that they can be input into the subsequent classification supervision part. The whole process can be expressed as:

[0023]

[0024]

[0025] in, represents the jth partition block, DP represents the dense partition operation, N represents the number of partition blocks, It represents the 1-dimensional compressed feature obtained by compressing the spatial dimension of the partition block using the average pooling layer, and its shape is 1024×1×1. In addition, j∈(0,10] represents the partition block obtained after the j-th dense strategy partitioning. The specific numbering order is as follows Figure 3 shown.

[0026] The specific structure of the classification supervision part: each compressed feature The custom classifier module is fed into the model to predict the geographic label of each compressed feature, and the cross entropy loss is used to calculate the difference between the output result and the true label. The algorithm is optimized by minimizing the cross entropy loss value to improve the accuracy of multi-view image matching.

[0027] The classifier module CF consists of a fully connected layer, a batch normalization layer, a dropout layer, and a classification layer (Cls). The fully connected layer has an input dimension of 1024 and an output dimension of 512; the batch normalization layer has an input dimension of 512; the classification layer has an input dimension of 512 and an output dimension of 701 (the total number of geo-tag categories).

[0028] The difference between the cross entropy loss calculation output and the true label is expressed as:

[0029]

[0030]

[0031] in, represents the predicted category output by the classifier module; Y represents the real geographic tag category corresponding to the input image, C represents the number of geographic tag categories, and CF represents the classifier module. In the present invention, the loss of each partition block is calculated separately.

[0032] Complete network model: The cross-view theorem positioning model based on the dense partitioning strategy proposed in this patent includes a drone branch and a satellite image branch, and the two branches adopt a shared weight learning method. Each branch consists of three parts: a feature extraction part, a feature partitioning processing part, and a classification supervision part. The initial input of the network is a given image X. According to the image's perspective type, it chooses to enter the drone branch or the satellite image branch; given the input image X, the test image is feature extracted through the SwinV2 backbone network to obtain high-dimensional features; the high-dimensional feature map is first partitioned using a dense partitioning strategy to obtain N partition blocks; then the partition blocks are averaged and pooled to obtain N 1-dimensional compressed features; the classifier module is used to predict the geographic label for each compressed feature, and the cross entropy loss is used to calculate the difference between the output result and the true label, thereby optimizing the network model.

[0033] Step 3: The trained geolocation network model receives the drone view image to be located and matches it with the corresponding satellite view image.

[0034] Load the dense partitioning network model weights trained in step 2 and update the parameters in the model. Secondly, pass the drone view image X to be located as input data into the network model, and the input data is sequentially processed through the feature extraction part and the feature partitioning part. In the classifier of the feature partitioning part, remove the classification layer Cls, compare the features of the front end of the classification layer Cls with the corresponding existing satellite image features, and use the Euclidean distance to measure the similarity of the two view images, that is, the correctly matched multi-view images have a shorter Euclidean distance, thereby screening out satellite images that match the drone view images. In addition, the network model uses the above model weights and does not need to be trained separately again. It can also handle drone navigation tasks, that is, one satellite image can be input, and the network outputs multiple drone images that match the satellite image.

[0035] The beneficial effects of the present invention are as follows:

[0036] The dense partitioning strategy proposed in this invention, on the one hand, divides the feature map into dense blocks to guide the network to learn local semantic features; on the other hand, the dense partitioning strategy retains the global features and can protect the global structure of the image from being destroyed. Therefore, the robustness of the algorithm to the size and position offset of objects in the image is enhanced from both the local and global levels, and the accuracy of multi-view image matching is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 : Schematic diagram of the method flow of the present invention;

[0038] Figure 2 : Flowchart of cross-perspective geolocation network based on dense partitioning strategy;

[0039] Figure 3 : Block strategy diagram for dense partitioning;

[0040] Figure 4 : Network model diagram using dense partitioning strategy;

[0041] Figure 5 : Visualization results;

[0042] Among them, Table 1 is the quantitative results on the public dataset University. DETAILED DESCRIPTION

[0043] The present invention will be described in detail below with reference to the accompanying drawings, and the purpose and effects of the present invention will become more apparent.

[0044] like Figure 1 and 2 As shown, the present invention proposes a cross-viewpoint geolocation method based on a dense partitioning strategy, and its basic steps are:

[0045] Step 1: Dataset creation

[0046] Step 1.1: Create two image datasets required for network training in step 2.

[0047] Download the public dataset University-1652 dataset. The original University-1652 dataset is a UAV-based geolocation dataset. It contains data from three platforms, namely, images of 1,652 university buildings from drone, satellite, and ground perspectives. The drone-view images contained in the dataset have less content occlusion, better visibility, and are aerial perspectives with satellite views. They have similar structures and are easier to match with satellite images. This dataset can be used for two tasks, namely, target positioning from the drone perspective and drone navigation. The training set includes 701 buildings from 33 universities, and the test set includes the other 951 buildings from the remaining 39 universities. There are no overlapping universities in the training and test sets. The drone and satellite perspectives in this data are used for drone view target positioning. There are 37,855 drone view images in the query set, 701 true matching satellite view images and 250 satellite view interference items in the library, and for each drone view image, there is only one true matching satellite view image. The goal of the drone geolocation task is to find the corresponding satellite view image based on the drone view image, so as to determine the geographic location of the drone.

[0048] Step 2: Construct a cross-view geolocation network model based on dense partitioning strategy

[0049] like Figure 4 As shown in the figure, the network model contains a drone branch and a satellite image branch, and the two branches adopt a shared weight learning method. Each branch consists of three parts: feature extraction part, feature division processing part and classification supervision part.

[0050] Specific structure of feature extraction part: This paper uses SwinV2 as the backbone network to extract visual features, and uses parameters trained based on the ImageNet-21K dataset as the pre-trained weights of the backbone network. Specifically, given an input image X with a size of 3×256×256, after the backbone network extracts features, we can obtain a high-dimensional feature map f with a shape of 1024×8×8 in each branch. i The corresponding operation can be expressed as

[0051] f i =F swinV2 (X i ),i∈{D,S}

[0052] The subscripts D and S represent the UAV and satellite perspectives, respectively; F swinV2 Represents the feature extraction operation of the backbone network.

[0053] Specific structure of feature partitioning processing part: This patent designs a dense partitioning strategy for high-dimensional feature graph f i Divide it into N (defined as N = 10) blocks with different shapes. First, the high-dimensional feature map f i Slice from the center to the outside into four non-overlapping annular areas, which are defined as basic partition blocks (P1, P2, P3, P4). The specific steps are as follows:

[0054] (a) Calculate the size increment Δh of each basic partition block,

[0055]

[0056] Among them, floor() means that the result is rounded down.

[0057] (b) For the high-dimensional feature map f i After partitioning, four basic partition blocks (P1, P2, P3, P4) are obtained. The partitioning order is as follows:

[0058]

[0059] The four basic partition blocks (P1, P2, P3, P4) obtained do not have overlapping areas. On this basis, by densely combining these basic partition blocks, 10 partition blocks with different shapes are obtained, such as Figure 3 As shown. Then the shape of each partition block can be expressed as follows

[0060]

[0061] Among them, ■ represents a square partition (solid); except for the square partition, other partitions are square ring partitions, that is, the difference between the large square partition and the small square partition (hollow).

[0062] like Figure 3 As shown in , the combined partition blocks can expand the boundaries of the basic partition blocks to varying degrees, thereby alleviating the problem of continuous features being divided into different regions. Then, the average pooling layer is used to compress the spatial dimensions of the partition blocks so that they can be input into the subsequent classification supervision part. The whole process can be described as

[0063]

[0064]

[0065] Where DP represents dense partitioning operation, It represents the 1D compressed feature obtained by compressing the spatial dimension of the partition block using the average pooling layer, and its shape is 1024×1×1. In addition, j∈(0,10] represents the jth dense strategy partition, and the numbering order is as follows Figure 3 shown.

[0066] The specific structure of the classification supervision part: each compressed feature The custom classifier module is fed into the model to predict the geographic label of each compressed feature, and the cross entropy loss is used to calculate the difference between the output result and the true label. The algorithm is optimized by minimizing the cross entropy loss value to improve the accuracy of multi-view image matching.

[0067] The classifier module includes a fully connected layer, a batch normalization layer, a dropout layer, and a classification layer (Cls). The fully connected layer has an input dimension of 1024 and an output dimension of 512; the batch normalization layer has an input dimension of 512; the classification layer has an input dimension of 512 and an output dimension of 701 (the total number of geo-tags).

[0068] The difference between the output of the cross entropy loss calculation and the true label can be expressed as:

[0069]

[0070]

[0071] Where C represents the number of geo-tag categories and CF represents the classifier module. Note that the loss for each partition block is calculated separately.

[0072] Complete network model: The cross-view theorem positioning model based on dense partitioning strategy proposed in this patent includes a drone branch and a satellite image branch, and the two branches adopt a shared weight learning method. Each branch consists of three parts: feature extraction part, feature partitioning processing part and classification supervision part. The initial input of the network is a given image X. According to the image's perspective type, it is selected to enter the drone branch or the satellite image branch; given image X, the test image is feature extracted through the SwinV2 backbone network to obtain high-dimensional features; the high-dimensional feature map is first divided into 10 partition blocks using a dense partitioning strategy; then the partition blocks are averaged and pooled to obtain 10 1-dimensional compressed features; the classifier module is used to predict the geographic label for each compressed feature, and the cross entropy loss is used to calculate the difference between the output result and the true label, thereby optimizing the network model.

[0073] Step 3: The trained neural network model receives the drone’s view image that needs to be located and matches it with the corresponding satellite image.

[0074] Load the dense partitioning network model weights trained in step 2 and update the parameters in the model. Secondly, pass the drone view image X that needs to be located into the network model as input data, and the input data is sequentially processed through the feature extraction part and the feature partitioning part. In the classifier of the feature partitioning part, remove the classification layer Cls, compare the obtained features with the existing satellite image features, and use the Euclidean distance to measure the similarity of the two view images, that is, the correctly matched multi-view images have a shorter Euclidean distance, thereby screening out satellite images that match the drone view images. In addition, the network model uses the above model weights and does not need to be trained separately again. It can also handle drone navigation tasks, that is, one satellite image can be input, and the network outputs multiple drone images that match the satellite image.

[0075] Furthermore, the present invention also provides a cross-viewpoint geographic positioning system based on a dense partitioning strategy, including a data acquisition module, a geographic positioning network model construction module and a testing module.

[0076] The data acquisition module is specifically implemented as follows: Download the public dataset University-1652 dataset, which is used for two tasks, namely target positioning from the drone perspective and drone navigation. The training set includes 701 buildings from 33 universities, and the test set includes the other 951 buildings from the remaining 39 universities, and there are no overlapping universities in the training set and the test set. The drone and satellite perspectives in the dataset are used for drone view target positioning. There are 37,855 drone view images in the query set, 701 true matching satellite view images and 250 satellite view interference items in the dataset, and for each drone view image, there is only one true matching satellite view image. The goal of the drone geolocation task is to find the corresponding satellite view image based on the drone view image, so as to determine the geographic location of the drone.

[0077] The geolocation network model building module includes a drone branch and a satellite image branch, and the two branches adopt a shared weight learning method. Each branch consists of three parts: a feature extraction part, a feature division processing part, and a classification supervision part.

[0078] Feature extraction part: SwinV2 is used as the backbone network to extract visual features, and the parameters trained based on the ImageNet-21K dataset are used as the pre-trained weights of the backbone network.

[0079] Feature partitioning processing part: Design dense partitioning strategy for high-dimensional feature graph f i Divide and obtain N (divided blocks with different shapes. First, the high-dimensional feature map f iSlice from the center to the outside into four non-overlapping annular areas, and define the annular areas as basic partition blocks P1, P2, P3 and P4 respectively. The specific steps are as follows:

[0080] (a) Calculate the size increment Δh of each basic partition block,

[0081]

[0082] Among them, floor() means that the result is rounded down.

[0083] (b) For the high-dimensional feature map f i After partitioning, four basic partition blocks (P1, P2, P3, P4) are obtained. The partitioning order is as follows:

[0084]

[0085] By densely combining these basic partition blocks, 10 partition blocks with different shapes are obtained, such as Figure 3 As shown. The shape of each partition block is expressed as follows:

[0086]

[0087] Among them, ■ represents a square partition (solid); except for the square partition, other partitions are square ring partitions, that is, the difference between the large square partition and the small square partition (hollow).

[0088] like Figure 3 As shown in Figure 2, the combined partition blocks can expand the boundaries of the basic partition blocks to varying degrees, thereby alleviating the problem of continuous features being divided into different regions. The average pooling layer is then used to compress the spatial dimensions of the partition blocks so that they can be input into the subsequent classification supervision part.

[0089] Classification supervision part: Each compressed feature is fed into a custom classifier module to predict the geographic label of each compressed feature, and the cross entropy loss is used to calculate the difference between the output result and the true label. The algorithm is optimized by minimizing the cross entropy loss value to improve the accuracy of multi-view image matching.

[0090] The classifier module CF consists of a fully connected layer, a batch normalization layer, a dropout layer, and a classification layer (Cls). The fully connected layer has an input dimension of 1024 and an output dimension of 512; the batch normalization layer has an input dimension of 512; the classification layer has an input dimension of 512 and an output dimension of 701 (the total number of geo-tag categories).

[0091] The test module is specifically implemented as follows: the trained geolocation network model receives the drone view image that needs to be located and matches it with the corresponding satellite view image.

[0092] Example:

[0093] In order to verify the effectiveness of the algorithm proposed in this invention, we use the test set of the public dataset University-1652 to evaluate the algorithm, including the UAV geolocation task (input UAV perspective image) and the UAV navigation task (input satellite perspective image). Table 1 shows the quantitative experimental results of using the dense partitioning strategy model proposed in this invention and the model without dense partitioning strategy. The dense partitioning strategy proposed in this invention can significantly improve the multi-view image matching accuracy in both tasks. In addition, Figure 5 The network visualization results are shown in the upper part of the figure: (1) Input a drone image and the model predicts the corresponding satellite image; the lower part of the figure: (2) Input a satellite image and the model predicts the corresponding drone image. The black boxes represent the correct prediction results, and the gray boxes represent the wrong prediction results.

[0094] Table 1 Quantitative results on the public dataset University

[0095]

[0096] In summary, from Figure 5 It can be clearly seen that the visualization results of the cross-view geolocation algorithm based on the dense partitioning strategy proposed by the present invention can accurately find images of other perspectives with the same geographic tag as the input image. Table 1 shows the quantitative results of the model based on the dense partitioning strategy proposed by the present invention and the model without the dense partitioning strategy. The dense partitioning strategy proposed in this patent densely blocks the high-dimensional feature map, guides the network model to mine local semantic features while achieving anti-interference to object size changes and position offsets, and significantly improves the multi-view image matching accuracy in two tasks. Specifically, compared with the basic model that does not use the dense partitioning strategy, the model using the dense partitioning strategy has achieved a performance improvement of 1.64% Recall@1 and 2.18% AP in the UAV positioning task (UAV-satellite), and a performance improvement of 3.09% Recall@1 and 3.11% AP in the UAV navigation task (satellite-UAV). Compared with the difficulty of improving the traditional accuracy, it can be clearly seen from the quantitative performance improvement data in Table 1 that the matching accuracy of the present invention has been significantly improved.

Claims

1. A cross-viewpoint geolocation method based on dense partitioning strategy, characterized in that The steps include: Step 1: Dataset creation: Create the image dataset required for network training; Step 2: Construct a cross-view geolocation network model based on dense partitioning strategy; The geolocation network model consists of a drone branch and a satellite image branch. The two branches adopt a shared weight learning method. Each branch consists of three parts: feature extraction part, feature partitioning processing part and classification supervision part. Feature extraction part: SwinV2 is used as the backbone network to extract visual features, and the parameters trained based on the ImageNet-21K dataset are used as the pre-trained weights of the backbone network; Feature partitioning processing part: Design dense partitioning strategy for high-dimensional feature graph f i Divide and obtain N divided blocks with different shapes: First, the high-dimensional feature map f i Slice from the center to the outside into four non-overlapping annular areas, and define the annular areas as basic partition blocks P1, P2, P3 and P4 respectively; secondly, densely combine the basic partition blocks to obtain N partition blocks with different shapes; then use the average pooling layer to compress the spatial dimensions of the partition blocks to obtain compressed features; Classification supervision part: each compressed feature is fed into the classifier module, the geographic label of each compressed feature is predicted, and the difference between the output result and the true label is calculated using the cross entropy loss; the algorithm is optimized by minimizing the cross entropy loss value to improve the accuracy of multi-view image matching; Step 3: The trained cross-view geolocation network model receives the drone view image to be located and matches it with the corresponding satellite image; The specific implementation of the feature segmentation processing part is as follows: (a) Calculate the size increment Δh of each basic partition block, Among them, floor() means that the result is rounded down; (b) For the high-dimensional feature map f i After partitioning, four basic partition blocks P1, P2, P3 and P4 are obtained. The partitioning order is as follows: The obtained four basic partition blocks P1, P2, P3 and P4 have no overlapping areas; The partition blocks of different shapes are expressed as follows: Among them, the ones with superscript ■ represent square partitions, and the other partitions without superscripts are square ring partitions, that is, the difference between the large square partition and the small square partition.

2. The cross-viewpoint geolocation method based on dense partitioning strategy according to claim 1, characterized in that: The feature extraction part is implemented as follows: Given an input image X with a size of 3×H×W, after the backbone network extracts features, a shape of The high-dimensional feature map f i , the corresponding operation is expressed as: f i =F swinV2 (X i ),i∈{D,S} (1) Wherein, the subscripts D and S represent the drone view and satellite view, respectively, H represents the height of the input image X, and w represents the width of the input image X; F swinV2 Represents the feature extraction operation of the backbone network.

3. The cross-viewpoint geolocation method based on dense partitioning strategy according to claim 1, characterized in that: The compression features are obtained as follows: in, represents the jth partition block, DP represents the dense partition operation, N represents the number of partition blocks, It represents the 1D compressed feature obtained by compressing the spatial dimension of the partition block using the average pooling layer, and its shape is 1024×1×1; j∈(0,10] represents the partition block obtained after the jth dense strategy partitioning.

4. The cross-viewpoint geolocation method based on dense partitioning strategy according to claim 1, characterized in that: The classifier module CF includes a fully connected layer, a batch normalization layer, a Dropout layer and a classification layer Cls, where the fully connected layer has an input dimension of 1024 and an output dimension of 512; the batch normalization layer has an input dimension of 512; the classification layer has an input dimension of 512 and an output dimension of 701, where the dimension of the classification layer output is the total number of categories of geo-tags.

5. The cross-viewpoint geolocation method based on dense partitioning strategy according to claim 4, characterized in that: The difference between the cross entropy loss calculation output and the true label is expressed as: in, Represents the predicted category output by the classifier module; Y represents the true geo-tag category corresponding to the input image, C represents the number of geo-tag categories, and CF represents the classifier module.

6. The cross-viewpoint geolocation method based on dense partitioning strategy according to claim 1, characterized in that: Step 3 is implemented as follows: First, the weights of the cross-view geolocation network model based on the dense partitioning strategy trained in step 2 are loaded, and the parameters in the model are updated; secondly, the drone view image to be located is passed into the model as the input image, and the input image is first processed by the feature extraction part and the feature partitioning part in sequence; In the classifier of the feature division processing part, the classification layer Cls is removed, and the features at the front end of the classification layer Cls are compared with the corresponding existing satellite image features. The Euclidean distance is used to measure the similarity of the two perspective images, that is, the correctly matched multi-perspective images have a shorter Euclidean distance, thereby screening out the satellite images that match the drone perspective images.

7. The cross-viewpoint geolocation method based on dense partitioning strategy according to claim 6, characterized in that: This network model does not need to be trained separately again and can also handle drone navigation tasks. That is, when one satellite image is input, the network model outputs multiple drone images that match the satellite image.

8. A cross-viewpoint geographic positioning system based on a dense partitioning strategy, using the cross-viewpoint geographic positioning method based on a dense partitioning strategy as described in any one of claims 1 to 7, characterized in that It includes data collection module, geolocation network model building module and testing module; The data collection module downloads the public data set University-1652 data set, the training set includes 701 buildings from 33 universities, the test set includes the other 951 buildings from the remaining 39 universities, and there are no overlapping universities in the training set and the test set; The geolocation network model building module includes a drone branch and a satellite image branch, and the two branches adopt a shared weight learning method; each branch consists of three parts: a feature extraction part, a feature division processing part, and a classification supervision part; Feature extraction part: SwinV2 is used as the backbone network to extract visual features, and the parameters trained based on the ImageNet-21K dataset are used as the pre-trained weights of the backbone network; Feature partitioning processing part: Design dense partitioning strategy for high-dimensional feature graph f i Divide and obtain N divided blocks with different shapes: First, the high-dimensional feature map f i Slice from the center to the outside into four non-overlapping annular areas, and define the annular areas as basic partition blocks P1, P2, P3 and P4 respectively; secondly, densely combine the basic partition blocks to obtain N partition blocks with different shapes; then use the average pooling layer to compress the spatial dimensions of the partition blocks to obtain compressed features; Classification supervision part: each compressed feature is fed into the classifier module, the geographic label of each compressed feature is predicted, and the difference between the output result and the true label is calculated using the cross entropy loss; the algorithm is optimized by minimizing the cross entropy loss value to improve the accuracy of multi-view image matching; In the test module, the trained geolocation network model receives the drone view image to be located and matches it with the corresponding satellite view image.

Images