Limited-view-field cross-view-angle geographic positioning method based on view enhancement contrast learning

Abstract

The invention discloses a limited view field cross-view-angle geographic positioning method based on view enhancement contrast learning, and belongs to the technical field of computer vision, deep learning and geographic information. The method comprises the following steps: executing view enhancement operation on an original panoramic ground view and a panoramic satellite view to generate a multi-view sample set; extracting each view feature embedding vector through a feature encoder sharing the weight; constructing a composite optimization target based on an InfoNCE loss function, and optimizing a total loss function through bidirectional cross-view-field contrast learning and in-view-field contrast learning to obtain an optimized feature encoder; and inputting a new limited field-of-view ground query image into the optimized encoder to extract features, performing similarity comparison with the satellite image feature library, and determining a geographic position. According to the method, the difference between the field of view and the viewpoint is reduced through active view enhancement, the feature robustness is enhanced through multi-target contrast learning, the positioning challenge caused by the limited field of view and unknown direction is effectively solved, the positioning precision and robustness are remarkably improved, and the method is suitable for multiple fields such as automatic driving and robot navigation.

Description

Technical Field

[0001] This invention belongs to the fields of computer vision, deep learning and geographic information technology, and specifically relates to a limited field of view cross-view geolocation method based on view-enhanced contrastive learning. Background Technology

[0002] Cross-field geolocation technology aims to determine the precise geographical location of a query image by matching it with a large-scale database of satellite (or aerial) imagery with geographic coordinates. This technology has broad application prospects in fields such as autonomous driving, robot navigation, emergency response, and smart tourism.

[0003] Most existing cross-field-of-view geolocation methods assume that the query image is a 360-degree panoramic image, in which case the ground image contains rich scene information. However, in practical applications, images are usually captured by ordinary devices such as smartphones and vehicle cameras. The images captured by these devices often have a limited field of view (e.g., 70 degrees or 90 degrees), and the shooting direction is random and unknown. This setting of "limited field of view and unknown direction" poses a severe challenge to geolocation tasks: 1. Significant difference in viewpoint: Ground images are taken from a horizontal perspective, while satellite images are taken from a vertical overhead perspective, resulting in a fundamental difference in viewpoint between the two.

[0004] 2. Asymmetric field of view information: Ground images with limited field of view contain only a small part of the information of the entire scene, while the matching satellite images are usually panoramic views covering the area, resulting in a serious asymmetry in information.

[0005] 3. Orientation uncertainty: The unknown shooting direction makes the correspondence between ground images and satellite images more ambiguous, making it difficult for the model to learn stable matching features.

[0006] To address these challenges, researchers have proposed several methods. For example, classic cross-view matching methods (such as CVM-Net and TransGeo) perform well on panoramic images, but their performance drops sharply when applied to limited field-of-view conditions. Dynamic similarity matching (DSM) methods estimate azimuth by dynamically calculating feature similarity across different directions; some methods combine global features from panoramic satellite images with local features from limited field-of-view regions; and others optimize from the perspective of loss functions, such as the ArcGeo loss function. However, most of these methods focus on learning shared features across different viewpoints, neglecting to fundamentally reduce the differences between ground views and satellite views in both the "field of view" and "viewpoint" dimensions. This neglect leads to significant difficulties for the model in learning robust feature representations, and localization performance still needs improvement.

[0007] Therefore, designing a geolocation method that can effectively bridge the viewpoint and field-of-view gap between finite field-of-view ground images and panoramic satellite images, and learn feature representations that are highly robust to changes in viewpoint and field of view, is a technical challenge that urgently needs to be solved in this field. Summary of the Invention

[0008] The technical problem to be solved by this invention is to address the shortcomings of the prior art by providing a limited field of view cross-view geolocation method based on view enhancement contrast learning. This method is used to solve the technical problem that, in practical applications, the limited field of view of the query image and the unknown shooting direction lead to huge differences in viewpoints between cross-view images, serious asymmetry of field of view information, and fuzzy directional relationships, which in turn cause the performance of existing geolocation methods to decline sharply.

[0009] The present invention adopts the following technical solution: A limited field-of-view cross-view geolocation method based on view-enhanced contrastive learning includes the following steps: S1. Perform view enhancement operations on the input original panoramic ground view and panoramic satellite view to generate a multi-view sample set containing a limited field of view ground view, a limited field of view satellite view, a bird's-eye view, and an enhanced satellite view; S2. Input all view images in the multi-view sample set obtained in step S1 into a feature encoder with shared weights to extract the feature embedding vectors corresponding to each view. S3. Using the feature embedding vectors of each view obtained in step S2 as input, construct a composite optimization objective based on the InfoNCE loss function, optimize the total loss function through bidirectional cross-view contrastive learning and intra-view contrastive learning, and obtain the optimized feature encoder. S4. Input the new limited field-of-view ground query image into the feature encoder optimized in step S3 to extract the feature embedding vector, compare it with the satellite image feature library for similarity, and determine the geographical location of the query image based on the ranking result.

[0010] Preferably, in step S1, the view enhancement operation includes: Symmetrical satellite view mask: Based on the limited field of view, a symmetrical sector mask region is generated on the panoramic satellite image to obtain a limited field of view satellite view; Ground image to bird's-eye view transformation: Transforming a ground image with a limited field of view into a bird's-eye view through geometric projection.

[0011] Preferably, the transformation from ground image to bird's-eye view includes: Define a blank bird's-eye view plane with the same size as a standard satellite image; assume the camera is located at the center of this plane and has a fixed virtual height h; for any pixel (x, y) on the bird's-eye view plane, calculate the pitch angle θ and azimuth angle α using trigonometric functions based on its relative position to the center point and the virtual height h; using the spherical projection model of the panoramic image, solve for its pixel coordinates (u, v) on the original panoramic ground image based on the calculated θ and α; assign the pixel value at coordinates (u, v) to the (x, y) point on the bird's-eye view plane to complete the mapping from the ground view to the bird's-eye view.

[0012] Preferably, in step S2, the feature encoder adopts the ConvNeXt-Base network model, and the dimension of the feature embedding vector is 512 to 2048.

[0013] Preferably, in step S3, the bidirectional cross-field contrast learning includes: Ground-to-satellite comparison: Using the feature embedding vector of the limited field of view ground view extracted in step S2 as the query vector, it is compared with the feature embedding vector of all satellite views extracted in step S2 in the same batch. The distance between the query vector and the corresponding positive sample satellite view feature embedding vector is reduced, while the distance between the query vector and the negative sample satellite view feature embedding vector is increased. Satellite-to-ground comparison: Using the feature embedding vector of the bird's-eye view extracted in step S2 as the query vector, it is compared with the feature embedding vectors of all limited field-of-view ground views extracted in step S2 within the same batch, thus narrowing the distance between the query vector and the feature embedding vector of the corresponding positive sample limited field-of-view ground view.

[0014] Preferably, in step S3, the in-field contrast learning includes: Ground field of view comparison: The feature embedding vector of the limited field of view ground view extracted in step S2 is compared with the feature embedding vector of the corresponding original panoramic ground view extracted in step S2 to narrow the distance between the two sets of vectors. Intra-satellite field-of-view comparison: The enhanced satellite view feature embedding vector extracted in step S2 is compared with the original panoramic satellite view feature embedding vector, the limited field-of-view satellite view feature embedding vector is compared with the original panoramic satellite view feature embedding vector, and the limited field-of-view satellite view feature embedding vector is compared with the bird's-eye view feature embedding vector to narrow the distance between each set of vectors.

[0015] Preferably, in step S3, the total loss function is a weighted sum of the basic contrastive learning loss, the bidirectional cross-field contrastive learning loss, and the intra-field contrastive learning loss; The basic contrastive learning loss is the loss calculated based on the InfoNCE loss function, which is the feature embedding vector of the original panoramic ground view extracted in step S2 and the feature embedding vector of the original panoramic satellite view. The network parameters of the feature encoder are updated by minimizing the total loss function through backpropagation to obtain the optimized feature encoder.

[0016] Preferably, the total loss function is:

[0017] in, It is a basic contrastive learning loss using panoramic images. arrive It is the weight hyperparameter for each loss.

[0018] Preferably, in step S4, the similarity comparison adopts the cosine similarity algorithm; the satellite image feature embedding vectors stored in the satellite image feature library are extracted from the panoramic satellite images in the database by the feature encoder optimized in step S3; the feature vectors in the satellite image feature library are sorted according to the similarity score, and the geographical location corresponding to the top-k satellite image with the highest ranking is taken as the positioning result of the new limited field of view ground query image.

[0019] Secondly, embodiments of the present invention provide a limited field-of-view cross-view geolocation system based on view-enhanced contrastive learning, comprising: The view enhancement module is used to perform view enhancement operations on the input original panoramic ground view and panoramic satellite view, and generate a multi-view sample set including a limited field of view ground view, a limited field of view satellite view, a bird's-eye view and an enhanced satellite view; The feature extraction module is used to input all view images in the multi-view sample set generated by the view enhancement module into the feature encoder with shared weights to extract the feature embedding vectors corresponding to each view. The contrastive learning optimization module is used to take the feature embedding vectors of each view obtained by the feature extraction module as input, construct a composite optimization objective based on the InfoNCE loss function, and optimize the total loss function through bidirectional cross-view contrastive learning and intra-view contrastive learning to obtain the optimized feature encoder. The positioning module is used to input the new limited field-of-view ground query image into the feature encoder optimized by the contrastive learning optimization module to extract the feature embedding vector, compare it with the satellite image feature library for similarity, and determine the geographical location of the query image based on the ranking result.

[0020] Thirdly, a computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the steps of the above-described limited field-of-view cross-view geolocation method based on view-enhanced contrastive learning.

[0021] Fourthly, embodiments of the present invention provide a computer-readable storage medium including a computer program that, when executed by a processor, implements the steps of the above-described limited field-of-view cross-view geolocation method based on view-enhanced contrastive learning.

[0022] Fifthly, a chip includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the steps of the above-described limited field-of-view cross-view geolocation method based on view-enhanced contrastive learning.

[0023] In a sixth aspect, embodiments of the present invention provide an electronic device including a computer program, which, when executed by the electronic device, implements the steps of the above-described limited field-of-view cross-view geolocation method based on view-enhanced contrastive learning.

[0024] Compared with the prior art, the present invention has at least the following beneficial effects: A limited field-of-view cross-view geolocation method based on view-enhanced contrastive learning is proposed. This method proactively generates a multi-view sample set, bridging the gap between terrestrial and satellite views. Combined with shared-weight feature extraction and multi-target contrastive learning, it ensures feature space consistency and robustness. The inference phase directly calls the optimized encoder to achieve end-to-end localization, resulting in a closed-loop and highly efficient process. This method fundamentally solves the problem of field-of-view and viewpoint differences between limited field-of-view and panoramic satellite imagery. By proactively reducing data-level differences through view enhancement strategies and strengthening cross-view consistency at the feature level through contrastive learning, the method significantly improves the robustness and localization accuracy of the model under limited field-of-view conditions.

[0025] Furthermore, symmetric masking operations alleviate problems such as asymmetric field of view information, and BEV transformation unifies the perspectives of ground and satellite images, reducing the difficulty of model learning from the data source and enhancing the comparability of features.

[0026] Furthermore, without relying on precise camera parameters or depth information, the viewpoint transformation is achieved solely through geometric relationships. This method is simple, efficient, and suitable for scenarios where parameters are unknown in practical applications.

[0027] Furthermore, by leveraging the powerful feature extraction capabilities of advanced backbone networks, we ensure that the extracted features have high expressiveness and discriminative power, thereby improving the overall performance of the model.

[0028] Furthermore, by comparing data from the ground to the satellite and from the satellite to the ground, the model is forced to learn bidirectional consistency features, thereby enhancing its ability to adapt to changes in perspective.

[0029] Furthermore, by comparing multiple sets of views within the field of view, the model's consistent understanding of features from different views (panoramic, partial, and enhanced) of the same location is improved, thereby enhancing the robustness of the features.

[0030] Furthermore, by comprehensively optimizing multiple comparison targets, the model is able to fully learn the core features of the geographical scene that remain unchanged under different view changes.

[0031] Furthermore, the explicit form of the loss function facilitates model optimization and reproduction, and the various weight hyperparameters allow for flexible adjustment of the importance of different comparison tasks.

[0032] Furthermore, cosine similarity can effectively measure the semantic similarity between feature vectors, and the Top-k mechanism improves the fault tolerance and practicality of localization.

[0033] It is understood that the beneficial effects of the second to sixth aspects mentioned above can be found in the relevant descriptions in the first aspect mentioned above, and will not be repeated here.

[0034] In summary, this invention effectively solves the problems of field of view, viewpoint differences, and orientation uncertainty in cross-view geolocation with limited field of view through innovative view enhancement and multiple contrast learning mechanisms, significantly improving positioning accuracy and robustness, and has important theoretical value and practical application prospects.

[0035] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0036] Figure 1 This is a flowchart illustrating the implementation of the present invention; Figure 2 A schematic diagram of the overall structure of a frame provided by the present invention; Figure 3 A schematic diagram of a computer device provided in an embodiment of the present invention; Figure 4 This is a block diagram of a chip provided according to an embodiment of the present invention.

[0037] Among them, 60. Computer equipment; 61. Processor; 62. Memory; 63. Computer program; 600. Electronic device; 610. Processing unit; 620. Storage unit; 6201. Random access memory unit; 6202. Cache memory unit; 6203. Read-only memory unit; 6204. Program / utility; 6205. Program module; 630. Bus; 640. Display unit; 650. Input / output interface; 660. Network adapter; 700. External device. Detailed Implementation

[0038] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0039] In the description of this invention, it should be understood that the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0040] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0041] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes such combinations. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. Additionally, the character " / " in this invention generally indicates that the preceding and following objects have an "or" relationship.

[0042] It should be understood that although terms such as first, second, third, etc., may be used in the embodiments of the present invention to describe the preset range, these preset ranges should not be limited to these terms. These terms are only used to distinguish the preset ranges from one another. For example, without departing from the scope of the embodiments of the present invention, the first preset range may also be referred to as the second preset range, and similarly, the second preset range may also be referred to as the first preset range.

[0043] Depending on the context, the word "if" as used here can be interpreted as "when," "when," "in response to determination," or "in response to detection." Similarly, depending on the context, the phrase "if determination" or "if detection (of the stated condition or event)" can be interpreted as "when determination," "in response to determination," "when detection (of the stated condition or event)," or "in response to detection (of the stated condition or event)."

[0044] The accompanying drawings illustrate various structural schematic diagrams according to embodiments disclosed in this invention. These drawings are not to scale, and some details have been enlarged for clarity, and some details may have been omitted. The shapes of the various regions and layers shown in the drawings, as well as their relative sizes and positional relationships, are merely exemplary and may deviate from reality due to manufacturing tolerances or technical limitations. Furthermore, those skilled in the art can design regions / layers with different shapes, sizes, and relative positions as needed.

[0045] This invention provides a limited field-of-view cross-view geolocation method based on view-enhanced contrastive learning. Instead of passively accepting huge differences in field of view and perspective, it actively generates a series of intermediate auxiliary views through a "view enhancement" strategy, building a "bridge" connecting ground and satellite views. On this basis, a multi-objective contrastive learning mechanism is designed to force the model to learn highly robust feature representations of the same geographical location that remain unchanged under various view changes from different dimensions. The view enhancement actively reduces the differences at the input data level, and the multi-objective contrastive learning strengthens the consistent expression of different viewpoints / fields of view of the same location in the feature space.

[0046] Please see Figure 1 This invention discloses a limited field-of-view cross-view geolocation method based on view-enhanced contrast learning, comprising the following steps: S1, View Enhancement To fundamentally reduce the discrepancy between ground views and satellite views, this invention designs two types of view enhancement strategies: Symmetric Satellite View Masking: Traditionally, a panoramic satellite image is directly matched with a ground image with a limited field of view. This invention generates a symmetrical, fan-shaped mask region on the panoramic satellite image based on the limited field of view (e.g., 90 degrees) of the query ground image, thus creating a "limited field of view satellite view." This new view forms a symmetrical relationship with the ground image in terms of field of view, greatly alleviating the problem of asymmetrical field of view information.

[0047] Transformation from Ground Image to Bird's-Eye View (BEV): To address the viewpoint discrepancy issue, this invention transforms a ground image with a limited field of view into a bird's-eye view (BEV) image through geometric projection. This BEV image maintains the same viewpoint as the satellite image, both being top-down perspectives. This transformation is based on the planar world assumption and geometric relationships, requiring no precise camera intrinsic and extrinsic parameters or depth information. The specific process is as follows: S101. Define a blank BEV plane with the same size as a standard satellite image. S102. Assume the camera is located at the center of the plane and has a fixed virtual height h; S103. For any pixel (x, y) on the BEV plane, the pitch angle θ and azimuth angle α of the point in three-dimensional space can be calculated using trigonometric functions based on its relative position to the center point and the virtual height h. S104. Using the spherical projection model of the panoramic image, the pixel coordinates (u, v) on the original panoramic ground image can be solved based on the calculated θ and α. S104. Assign the pixel value at coordinates (u, v) to the point (x, y) on the BEV plane. By traversing all the pixels on the BEV plane, complete the dense mapping from the ground view to the BEV view. Since the input is a ground image with a limited field of view, the transformed BEV image only contains information within that field of view.

[0048] Through the two view enhancement operations described above, multiple new views, such as (limited field of view satellite image, ground BEV image), are generated for each pair of original (limited field of view ground image, panoramic satellite image) samples.

[0049] S2, Feature Extraction with Shared Weights This invention employs a Siamese network-like architecture, but all branches share the same deep convolutional neural network as the feature extractor (encoder). In this example, the ConvNeXt-Base network model is chosen as the backbone network due to its powerful feature representation capabilities. All images generated after view augmentation (such as finite field-of-view ground maps) BEV diagram Limited field-of-view satellite image The original panoramic images (q and r) are input into this shared feature encoder to extract high-dimensional feature embedding vectors, outputting global embedding features with dimensions ranging from 512 to 2048, preferably 1024 dimensions. The shared weight mechanism ensures that features from all different views reside in the same unified and comparable feature space.

[0050] S3, Multi-objective Comparative Learning Optimization To enable the model to learn robust and discriminative features, this invention designs a composite optimization objective that includes bidirectional cross-view contrastive learning and intra-view contrastive learning. All contrastive learning losses adopt the standard InfoNCE loss function form:

[0051] Where q is the query feature, R is the set of candidate features, r+ is the positive sample of q, and τ is the learnable temperature hyperparameter.

[0052] S301. Bidirectional Cross-View Contrastive Learning: Ground to Satellite (Q R): Features of a finite field-of-view ground map As a query, it is compared with the feature R of all satellite images in the same batch. The goal is to narrow the distance between it and the corresponding positive sample satellite image feature r, and widen the distance between it and the features of other negative sample satellite images. The first contrastive loss is denoted as . Using limited field-of-view ground image features For the query, the corresponding panoramic satellite image feature R is used as a positive sample for optimization.

[0053] Satellite (BEV) to Ground (R) Q): Features of BEV maps obtained by transforming ground maps As a query, it is compared with the features of all finite field-of-view ground maps within the same batch. The comparison aims to narrow the gap between the original sample and its corresponding positive sample. The distance. The second contrast loss is denoted as... using bird's-eye view image features For querying, use the corresponding finite field-of-view ground image features. Optimize for positive samples.

[0054] This bidirectional learning mechanism forces the model to not only identify satellite scenes from a ground perspective, but also to identify ground scenes from a satellite-like perspective (BEV), which greatly enhances the robustness and consistency of cross-field features.

[0055] S302, Intra-View Contrastive Learning: Intra-q: Features of a finite field-of-view ground map The model learns by contrasting its corresponding panoramic ground map features q. The goal is to enable the model to learn features representative of the entire panoramic scene from a limited local perspective, thereby enhancing its "perception" and "imagination" of the surrounding environment. The third contrastive loss is denoted as... Used to bring ground image features closer together in a limited field of view The distance between it and its corresponding panoramic ground image feature Q.

[0056] Intra-r (Satellite Field of View): Similarly, multiple comparisons are performed from the satellite's perspective. For example, features of satellite images that have undergone data augmentation (such as random cropping, rotation, etc.) are compared. Compare with the original panoramic satellite image feature R (fourth contrast loss) Used to zoom in on enhanced satellite image features (Distance between the original panoramic satellite image features R); finite field-of-view satellite image features generated by masking. Compare with panoramic satellite image feature R (fifth contrast loss) Features of limited field-of-view satellite images used for zooming in on mask generation (distance between the original panoramic satellite image features R); and the limited field-of-view satellite image features Features of ground-based BEV maps Compare (Sixth comparison loss) Features of limited field-of-view satellite images used for zooming in on mask generation Corresponding bird's-eye view image features (Distance between them). These objectives collectively improve the model's consistency in representing features of the same location across different satellite views (panoramic, partial, enhanced) and cross-modal overhead views (satellite image, BEV image).

[0057] The final total loss function is the weighted sum of all the contrastive learning loss terms mentioned above:

[0058] in, It is a basic contrastive learning loss using panoramic images. arrive It is the weight hyperparameter for each loss.

[0059] By optimizing this total loss function, the model can comprehensively learn the core features of the geographical scene that remain unchanged under drastic changes in viewpoint and field of view.

[0060] S4, Reasoning and Positioning After model training is complete, the inference phase begins. For a new ground query image with a limited field of view, the following steps are performed: S401. Input it into the trained shared feature encoder and extract its feature embedding vector; S402. Compare the query vector with a pre-calculated feature library containing all satellite image features in the database (usually using cosine similarity). S403. Sort the satellite images in the database according to the similarity score; S404 The geographical location corresponding to the highest-ranked (Top-k) satellite image is the location result for the query image.

[0061] Please see Figure 2 In another embodiment of the present invention, a limited field-of-view cross-view geolocation system based on view enhancement contrastive learning is provided. This system can be used to implement the above-mentioned limited field-of-view cross-view geolocation method based on view enhancement contrastive learning. Specifically, the limited field-of-view cross-view geolocation system based on view enhancement contrastive learning includes a view enhancement module, a feature extraction module, and a contrastive learning optimization module.

[0062] The view enhancement module receives the input raw image pair, namely the panoramic ground view (q) and the panoramic satellite view (r), and generates a series of enhanced views.

[0063] First, a finite field-of-view ground view (q_fov) is cropped from the panoramic ground view (q) according to a preset finite field-of-view angle (e.g., 90°) and a random direction. This simulates the query image in a real-world application.

[0064] View Mask operation: Receives panoramic satellite view (r) and finite field of view information, and generates a finite field of view satellite view (r_fov). This operation ensures that r_fov is symmetrical to q_fov in terms of the field of view.

[0065] BEV Transformation operation: Receive a finite field-of-view ground view (q_fov) and generate a bird's-eye view (r_bev) using the aforementioned geometric projection method.

[0066] Meanwhile, in order to perform in-field contrast learning, standard data augmentation (such as random cropping, flipping, etc.) is performed on the panoramic satellite view (r) to obtain the augmented satellite view (r_aug).

[0067] After processing by this module, an original sample pair (q, r) is expanded into a sample set containing multiple views: {q, q_fov, r, r_fov, r_bev, r_aug}.

[0068] Feature extraction module It consists of a shared-weight encoder (Encoder-q, Encoder-r). Figure 2 In this context, all the squares marked "Encoder" represent the same network instance, and they share the exact same set of network parameters.

[0069] All view images from the above sample set are fed into the encoder. The encoder extracts deep features from each input image and outputs a fixed-dimensional feature embedding vector. The output vector dimension is set to 1024. This yields the feature vectors corresponding to each view: f_q, f_q_fov, f_r, f_r_fov, f_r_bev, f_r_aug.

[0070] Comparative learning optimization module A multi-group contrastive learning task based on the InfoNCE loss function. Figure 2 In the diagram, the lines connecting different feature vectors and the words "InfoNCE Loss" represent these contrast relationships. f_q is compared with f_r as the base for cross-field contrast.

[0071] Two-way cross-field comparison: The f_q_fov (query) is compared with all f_r (positive and negative samples) in the batch.

[0072] The f_r_bev (query) is compared with all f_q_fov (positive and negative samples) in the batch.

[0073] In-field contrast: The f_q_fov (query) is compared with all f_q (positive and negative samples) in the batch.

[0074] The f_r_fov (query) is compared with all f_r (positive and negative samples) in the batch.

[0075] The f_r_aug (query) is compared with all f_r (positive and negative samples) in the batch.

[0076] The f_r_fov (query) is compared with all f_r_bev (positive and negative samples) in the batch.

[0077] The positioning module is used to input the new limited field-of-view ground query image into the feature encoder optimized by the contrastive learning optimization module to extract the feature embedding vector, compare it with the satellite image feature library for similarity, and determine the geographical location of the query image based on the ranking result.

[0078] This invention provides a terminal device comprising a processor and a memory. The memory stores a computer program, which includes program instructions. The processor executes the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, graphics processing units (GPUs), tensor processing units (TPUs), digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions to achieve a corresponding method flow or function. The processor described in this embodiment can be used for the operation of a limited field-of-view cross-view geolocation method based on view-enhanced contrastive learning, including: View augmentation operations are performed on the input original panoramic ground view and panoramic satellite view to generate a multi-view sample set containing a limited field-of-view ground view, a limited field-of-view satellite view, a bird's-eye view, and an augmented satellite view. All view images in the obtained multi-view sample set are input into a feature encoder with shared weights to extract the feature embedding vectors corresponding to each view. Using the obtained feature embedding vectors of each view as input, a composite optimization objective is constructed based on the InfoNCE loss function. The total loss function is optimized through bidirectional cross-field-of-view contrastive learning and intra-field-of-view contrastive learning to obtain the optimized feature encoder. A new limited field-of-view ground query image is input into the optimized feature encoder to extract its feature embedding vector, and a similarity comparison is performed with the satellite image feature library. The geographical location of the query image is determined based on the ranking result.

[0079] Please see Figure 3The terminal device is a computer device. In this embodiment, the computer device 60 includes a processor 61, a memory 62, and a computer program 63 stored in the memory 62 and executable on the processor 61. When executed by the processor 61, the computer program 63 implements the limited field-of-view cross-view geolocation method based on view-enhanced contrastive learning in this embodiment. To avoid repetition, these details are not elaborated here. Alternatively, when executed by the processor 61, the computer program 63 implements the functions of each model / unit in the limited field-of-view cross-view geolocation system based on view-enhanced contrastive learning in this embodiment. To avoid repetition, these details are not elaborated here.

[0080] Computer device 60 can be a desktop computer, laptop, handheld computer, cloud server, or other computing device. Computer device 60 may include, but is not limited to, a processor 61 and a memory 62. Those skilled in the art will understand that... Figure 3 This is merely an example of computer device 60 and does not constitute a limitation on computer device 60. It may include more or fewer components than shown, or combine certain components, or different components. For example, computer device may also include input / output devices, network access devices, buses, etc.

[0081] The processor 61 may be a Central Processing Unit (CPU), or other general-purpose processors, graphics processing units (GPUs), tensor processing units (TPUs), digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0082] The memory 62 can be an internal storage unit of the computer device 60, such as a hard disk or RAM of the computer device 60. The memory 62 can also be an external storage device of the computer device 60, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., provided on the computer device 60.

[0083] Furthermore, the memory 62 may include both internal storage units of the computer device 60 and external storage devices. The memory 62 is used to store computer programs and other programs and data required by the computer device. The memory 62 can also be used to temporarily store data that has been output or will be output.

[0084] Please see Figure 4 The terminal device is an electronic device 600, which is manifested in the form of a general-purpose computing device. The components of the electronic device may include, but are not limited to: at least one processing unit 610, at least one storage unit 620, a bus 630 connecting different platform components (including storage unit 620 and processing unit 610), a display unit 640, etc.

[0085] The storage unit stores program code, which can be executed by the processing unit 610 to perform the steps described in the method section of this specification according to various exemplary embodiments of the present invention. For example, the processing unit 610 can perform actions such as... Figure 1 The steps are shown in the figure.

[0086] Storage unit 620 may include a readable medium in the form of a volatile storage unit, such as random access memory (RAM) 6201 and / or cache memory 6202, and may further include a read-only memory (ROM) 6203.

[0087] Storage unit 620 may also include a program / utility 6204 having a set (at least one) program module 6205, such program module 6205 including but not limited to: operating system, one or more application programs, other program modules and program data, each or some combination of these examples may include an implementation of a network environment.

[0088] Bus 630 can represent one or more of several types of bus structures, including a memory cell bus or memory cell controller, a peripheral bus, a graphics acceleration port, a processing unit, or a local bus using any of the multiple bus structures.

[0089] Electronic device 600 can also communicate with one or more external devices 700 (e.g., keyboard, pointing device, Bluetooth device, etc.), and with one or more devices that enable a user to interact with electronic device 600, and / or with any device that enables electronic device 600 to communicate with one or more other computing devices (e.g., router, modem). This communication can be performed via input / output interface 650. Furthermore, electronic device 600 can also communicate with one or more networks (e.g., local area network, wide area network, and / or public network, such as the Internet) via network adapter 660. Network adapter 660 can communicate with other modules of electronic device 600 via bus 630. It should be understood that, although not shown in the figures, other hardware and / or software modules can be used in conjunction with electronic device 600, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage platforms.

[0090] Example 4 This invention also provides a storage medium, specifically a computer-readable storage medium, which is a memory device in a terminal device for storing programs and data. It is understood that the computer-readable storage medium here can include both built-in storage media in the terminal device and extended storage media supported by the terminal device; it can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. The computer-readable storage medium provides storage space that stores the terminal's operating system. Furthermore, the storage space also stores one or more instructions suitable for loading and execution by a processor, which can be one or more computer programs (including program code). More specific examples of the computer-readable storage medium include: an electrical connection with one or more wires, a portable disk, a hard disk, random access memory, read-only memory, erasable programmable read-only memory, optical fiber, portable compact disk read-only memory, optical storage device, magnetic storage device, or any suitable combination thereof.

[0091] Computer-readable storage media also include data signals propagated in baseband or as part of a carrier wave, carrying readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable storage medium can also be any readable medium other than a readable storage medium that can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the readable storage medium can be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, radio frequency, etc., or any suitable combination thereof.

[0092] Program code for performing the operations of this invention can be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java and C++, and conventional procedural programming languages ​​such as C or similar languages. The program code can execute entirely on the user's computing device, partially on the user's computing device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0093] One or more instructions stored in a computer-readable storage medium can be loaded and executed by a processor to implement the corresponding steps of the limited field-of-view cross-view geolocation method based on view-enhanced contrast learning in the above embodiments; one or more instructions in the computer-readable storage medium are loaded by the processor and executed as follows: View augmentation operations are performed on the input original panoramic ground view and panoramic satellite view to generate a multi-view sample set containing a limited field-of-view ground view, a limited field-of-view satellite view, a bird's-eye view, and an augmented satellite view. All view images in the obtained multi-view sample set are input into a feature encoder with shared weights to extract the feature embedding vectors corresponding to each view. Using the obtained feature embedding vectors of each view as input, a composite optimization objective is constructed based on the InfoNCE loss function. The total loss function is optimized through bidirectional cross-field-of-view contrastive learning and intra-field-of-view contrastive learning to obtain the optimized feature encoder. A new limited field-of-view ground query image is input into the optimized feature encoder to extract its feature embedding vector, and a similarity comparison is performed with the satellite image feature library. The geographical location of the query image is determined based on the ranking result.

[0094] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0095] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0096] Experimental verification To fully verify the effectiveness and superiority of the method of this invention, systematic and detailed experiments were conducted on the publicly available large-scale cross-field geolocation dataset CVUSA. This dataset contains more than 35,000 pairs of panoramic ground images and corresponding satellite images, covering various geographical scenes such as cities, suburbs, and rural areas. The images have uniform resolution and are accurately labeled with geographic coordinates, making it an authoritative benchmark in the field of cross-field geolocation.

[0097] The experimental environment was configured with an Intel Core i9-12900K processor, an NVIDIA RTX 3090 graphics card, and 24GB of RAM. The software framework used was PyTorch 1.10. The batch size was set to 32 during training, the initial learning rate was 0.0001, and the learning rate was adjusted using a cosine annealing strategy. The total number of training rounds was 60. The weight hyperparameters from \({γ}_{1}\) to \({γ}_{6}\) were set to 1.0, 1.0, 1.0, 1.0, 1.0, and 1.0, respectively, and the temperature hyperparameter τ was initialized to 0.07.

[0098] The comparative experiments selected mainstream state-of-the-art methods in the field, including the classic CVM-Net and TransGeo, the dynamic similarity matching method DSM, W2W-BEV combined with BEV transform, and ArcGeo with optimized loss function. The evaluation metric used was the industry-standard Top-k recall (R@k), including R@1, R@5, R@10, and R@1%, comprehensively measuring the model's localization performance at different ranking thresholds.

[0099] Experimental results show that, with a field of view of 180° (panoramic reference scene), the method of this invention achieves R@1 of 91.47% and R@5 of 97.50%, significantly better than W2W-BEV's 86.55% and 95.63%, respectively. In the more challenging 90° field of view scene, the method of this invention achieves R@1 of 68.54% and R@10 of 91.37%, far exceeding W2W-BEV's 64.75% and 89.37%, respectively. In the 70° narrow field of view scene, the method of this invention still maintains R@1 of 56.84% and R@5 of 78.58%, while the best-performing method among the comparisons, W2W-BEV, only achieves 43.33% and 68.73%, fully demonstrating the overwhelming advantage of this invention in limited field of view scenarios.

[0100] To further verify the contributions of each core module, ablation experiments were conducted. When the view augmentation strategy (w / oVACL) was removed, R@1 decreased to 55.50% and R@10 decreased to 81.50% at a 90° field of view; this clearly demonstrates that view augmentation and multi-object contrast learning are key to performance improvement. Furthermore, stability tests show that the performance fluctuation of this invention in limited field-of-view image localization under different shooting directions and lighting conditions is less than 3%, exhibiting extremely strong robustness.

[0101] Table 1 shows a comparison of the experimental performance of the present invention with existing methods.

[0102] Table 1 presents the performance comparison results of our proposed method (Ours) with six existing state-of-the-art methods on the publicly available large-scale cross-view geolocation dataset CVUSA. The core focus is on the "Top-k recall rate (R@k) under different field of view (FoV)". The table is divided horizontally into three field of view scenarios (FoV=180°, 90°, 70°), with four evaluation metrics (R@1, R@5, R@10, R@1%) set for each scenario. Vertically, it represents the comparison methods, covering representative solutions from different technical routes such as classic cross-view matching, dynamic similarity optimization, BEV transform fusion, and loss function improvement, fully covering the current mainstream technical level in the field.

[0103] Table 2 shows the ablation experimental results of the present invention.

[0104] Table 2 shows the ablation experiment results that verify the contribution of the core module of this invention (view enhancement strategy, denoted as VCL). Focusing only on the most challenging FoV=90° scene, the four evaluation indicators (R@1, R@5, R@10, R@1%) of the two methods are compared with "with or without view enhancement strategy" as the only variable, directly quantifying the performance improvement effect of view enhancement strategy.

[0105] Based on the data in Table 2, it is clear that the View Enhancement Strategy (VACL) is one of the core pillars that allows this invention to outperform existing technologies. Its essence is "actively reducing differences from the data preprocessing level"—existing methods passively accept differences in field of view / viewpoint, compensating only through feature learning; while this invention, through view enhancement, first transforms the "cross-dimensional difference" between "limited field of view ground map - panoramic satellite image" into "low-dimensional difference" between "limited field of view ground map - limited field of view satellite image" and "BEV map - satellite image," and then strengthens feature consistency through comparative learning, forming a dual advantage of "data optimization + model optimization," ultimately achieving a performance breakthrough.

[0106] This invention actively reduces the differences in field of view and viewpoint between finite field-of-view terrestrial images and panoramic satellite images through view enhancement strategies (symmetric satellite view masking and BEV transformation of ground images), constructing a "bridge" connecting the two perspectives. Based on this, a designed multi-objective contrastive learning mechanism (including bidirectional cross-field-of-view contrast and multi-field-of-view intra-view contrast) forces the model to learn robust feature representations of the same geographic location that remain unchanged under various view variations. Experimental results on the public dataset CVUSA show that, under finite field-of-view conditions (e.g., 70°, 90°), this invention significantly outperforms existing mainstream methods in various evaluation metrics (e.g., R@1), especially in more challenging small field-of-view scenarios. Ablation experiments further verify the key role of view enhancement and contrastive learning design in performance improvement. This invention provides an effective technical solution for achieving high-precision and highly robust cross-field-of-view geolocation.

[0107] In summary, this invention presents a limited field-of-view cross-view geolocation method based on view-enhanced contrastive learning. Through innovative data preprocessing and model optimization strategies, it successfully addresses the challenges posed by limited field of view and unknown orientation, providing an effective technical solution for achieving high-precision and robust cross-field-of-view geolocation.

[0108] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0109] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0110] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed in this invention can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0111] In the embodiments provided by this invention, it should be understood that the disclosed devices / terminals and methods can be implemented in other ways. For example, the device / terminal embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0112] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0113] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0114] If the integrated module / unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random-access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0115] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus, and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0116] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0117] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0118] The above content is only for illustrating the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solution based on the technical concept proposed in this invention shall fall within the scope of protection of the claims of this invention.

Claims

1. A limited field-of-view cross-view geolocation method based on view-enhanced contrastive learning, characterized in that, Includes the following steps: S1. Perform view enhancement operations on the input original panoramic ground view and panoramic satellite view to generate a multi-view sample set containing a limited field of view ground view, a limited field of view satellite view, a bird's-eye view, and an enhanced satellite view; S2. Input all view images in the multi-view sample set obtained in step S1 into a feature encoder with shared weights to extract the feature embedding vectors corresponding to each view. S3. Using the feature embedding vectors of each view obtained in step S2 as input, construct a composite optimization objective based on the InfoNCE loss function, optimize the total loss function through bidirectional cross-view contrastive learning and intra-view contrastive learning, and obtain the optimized feature encoder. S4. Input the new limited field-of-view ground query image into the feature encoder optimized in step S3 to extract the feature embedding vector, compare it with the satellite image feature library for similarity, and determine the geographical location of the query image based on the ranking result.

2. The limited field-of-view cross-view geolocation method based on view-enhanced contrastive learning according to claim 1, characterized in that, In step S1, the view enhancement operation includes: Symmetrical satellite view mask: Based on the limited field of view, a symmetrical sector mask region is generated on the panoramic satellite image to obtain a limited field of view satellite view; Ground image to bird's-eye view transformation: Transforming a ground image with a limited field of view into a bird's-eye view through geometric projection.

3. The limited field-of-view cross-view geolocation method based on view-enhanced contrastive learning according to claim 2, characterized in that, The transformation from ground image to bird's-eye view includes: Define a blank bird's-eye view plane with the same size as a standard satellite image; assume the camera is located at the center of this plane and has a fixed virtual height h; for any pixel (x, y) on the bird's-eye view plane, calculate the pitch angle θ and azimuth angle α using trigonometric functions based on its relative position to the center point and the virtual height h; using the spherical projection model of the panoramic image, solve for its pixel coordinates (u, v) on the original panoramic ground image based on the calculated θ and α; assign the pixel value at coordinates (u, v) to the (x, y) point on the bird's-eye view plane to complete the mapping from the ground view to the bird's-eye view.

4. The limited field-of-view cross-view geolocation method based on view-enhanced contrastive learning according to claim 1, characterized in that, In step S2, the feature encoder uses the ConvNeXt-Base network model, and the dimension of the feature embedding vector is 512 to 2048.

5. The limited field-of-view cross-view geolocation method based on view-enhanced contrastive learning according to claim 1, characterized in that, In step S3, bidirectional cross-field contrast learning includes: Ground-to-satellite comparison: Using the feature embedding vector of the limited field of view ground view extracted in step S2 as the query vector, it is compared with the feature embedding vector of all satellite views extracted in step S2 in the same batch. The distance between the query vector and the corresponding positive sample satellite view feature embedding vector is reduced, while the distance between the query vector and the negative sample satellite view feature embedding vector is increased. Satellite-to-ground comparison: Using the feature embedding vector of the bird's-eye view extracted in step S2 as the query vector, it is compared with the feature embedding vectors of all limited field-of-view ground views extracted in step S2 within the same batch, thus narrowing the distance between the query vector and the feature embedding vector of the corresponding positive sample limited field-of-view ground view.

6. The limited field-of-view cross-view geolocation method based on view-enhanced contrastive learning according to claim 1, characterized in that, In step S3, the in-field contrast learning includes: Ground field of view comparison: The feature embedding vector of the limited field of view ground view extracted in step S2 is compared with the feature embedding vector of the corresponding original panoramic ground view extracted in step S2 to narrow the distance between the two sets of vectors. Intra-satellite field-of-view comparison: The enhanced satellite view feature embedding vector extracted in step S2 is compared with the original panoramic satellite view feature embedding vector, the limited field-of-view satellite view feature embedding vector is compared with the original panoramic satellite view feature embedding vector, and the limited field-of-view satellite view feature embedding vector is compared with the bird's-eye view feature embedding vector to narrow the distance between each set of vectors.

7. The limited field-of-view cross-view geolocation method based on view-enhanced contrastive learning according to claim 1, characterized in that, In step S3, the total loss function is a weighted sum of the basic contrastive learning loss, the bidirectional cross-field contrastive learning loss, and the intra-field contrastive learning loss; The basic contrastive learning loss is the loss calculated based on the InfoNCE loss function, which is the original panoramic ground view feature embedding vector extracted in step S2 and the original panoramic satellite view feature embedding vector. The optimized feature encoder is obtained by minimizing the total loss function through backpropagation and updating the network parameters of the feature encoder.

8. The limited field-of-view cross-view geolocation method based on view-enhanced contrastive learning according to claim 7, characterized in that, The total loss function is: in, It is a basic contrastive learning loss using panoramic images. arrive It is the weight hyperparameter for each loss.

9. The limited field-of-view cross-view geolocation method based on view-enhanced contrastive learning according to claim 1, characterized in that, In step S4, the similarity comparison uses the cosine similarity algorithm; the satellite image feature embedding vectors stored in the satellite image feature library are extracted from the panoramic satellite images in the database by the feature encoder optimized in step S3; the feature vectors in the satellite image feature library are sorted according to the similarity score, and the geographical location corresponding to the top-k satellite image with the highest ranking is taken as the positioning result of the new limited field of view ground query image.

10. A limited field-of-view cross-view geolocation system based on view-enhanced contrastive learning, characterized in that, include: The view enhancement module is used to perform view enhancement operations on the input original panoramic ground view and panoramic satellite view, and generate a multi-view sample set including a limited field of view ground view, a limited field of view satellite view, a bird's-eye view and an enhanced satellite view; The feature extraction module is used to input all view images in the multi-view sample set generated by the view enhancement module into the feature encoder with shared weights to extract the feature embedding vectors corresponding to each view. The contrastive learning optimization module is used to take the feature embedding vectors of each view obtained by the feature extraction module as input, construct a composite optimization objective based on the InfoNCE loss function, and optimize the total loss function through bidirectional cross-view contrastive learning and intra-view contrastive learning to obtain the optimized feature encoder. The positioning module is used to input the new limited field-of-view ground query image into the feature encoder optimized by the contrastive learning optimization module to extract the feature embedding vector, compare it with the satellite image feature library for similarity, and determine the geographical location of the query image based on the ranking result.

Images