Cross-view geographic positioning acceleration method based on semantic description

Abstract

The invention relates to the technical field of computer vision and geographic information, and particularly discloses a cross-view geographic positioning acceleration method based on semantic description, which comprises the following steps of: performing instruction fine tuning on an input query image and a database image by adopting a visual language model to generate natural language description; coding the natural language description into a low-dimensional semantic vector by utilizing a word embedding model; constructing a semantic keyword association graph based on a graph embedding technology, and fusing a general semantic knowledge base and a geographic domain ontology to construct a surface feature special semantic tree; screening candidate subsets according to a preset ground feature priority and a saliency detection result of the query image; extracting global features and local key point descriptors of the images in the candidate subsets by adopting a lightweight model; constructing a visual similarity calculation model based on the state space model, and modeling a feature sequence dependency relationship; and fusing the semantic matching score and the visual similarity, outputting a final matching result by adopting a weighted sorting strategy, and cooperatively accelerating according to the matching result in combination with a lightweight model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer vision and geographic information, and particularly relates to a cross-view geographic positioning acceleration method based on semantic description. BACKGROUND

[0002] Current cross-view geographic positioning technology mainly includes two types, traditional methods based on bottom feature matching and end-to-end methods based on deep learning. The principle of the traditional method based on bottom feature matching is to extract local features (such as SIFT and SURF feature points) or global features (such as HOG and GIST) of an image, calculate feature similarity for matching. For example, thousands of feature points are extracted from satellite images and ground images, and after matching through the Euclidean distance, the RANSAC algorithm is used to remove the mis-matching points, and finally the coordinate mapping is calculated through the homography matrix. However, the traditional method has the disadvantages of high computational complexity, poor cross-view robustness, and dependence on high-resolution data. For example, for a satellite image library of 100,000 images, a single matching needs to traverse tens of thousands of images, and the time complexity is O (N 2 ), and the processing time of a single frame is more than 500 ms, which cannot meet the real-time requirements of automatic driving (≤100 ms / frame). The large sky-ground view angle difference will cause the same ground object to present significant geometric variation in the image, and the bottom features are prone to mis-matching. When facing light changes (such as day and night, seasonal differences), the pixel-level gray value will change dramatically, and the feature descriptor will fail. The deep learning method has improved performance compared with the traditional method, and the core principle is to use a convolutional neural network (CNN) or a Transformer to directly learn the cross-view mapping relationship between satellite images and ground images. For example, a twin network structure is used to input two types of images into an encoder with shared weights to extract features, and the cosine similarity is used to measure the matching; or a generative adversarial network (GAN) is used to generate cross-view images to assist in training the positioning model.

[0003] Although the end-to-end method based on deep learning has improved the efficiency of processing images and the accuracy of matching, it still has problems such as insufficient semantic representation, weak generalization of dynamic scenes, and large consumption of computing resources. The existing model relies on pixel-level or shallow semantic features (such as edges and textures), lacks understanding of high-level semantics such as “road type” and “landmark building”, and is difficult to distinguish similar scenes (such as two-way six-lane roads in different cities). In the face of dynamic occlusion (such as moving vehicles blocking buildings) or unstructured environments (such as temporary construction areas), the model's generalization ability decreases significantly, and a large amount of labeled data is needed for retraining, which is high in adaptation cost. And the typical Transformer model has more than 100M parameters and hundreds of billions of FLOPs, which requires expensive computing chips for running on embedded devices, limiting its practical application. SUMMARY

[0004] To achieve the object of the present application, the application provides a cross-view geographic positioning acceleration method based on semantic description, comprising:

[0005] Step S1: using BLIP-2 visual language model to fine-tune the instructions for the input query image and the database image, and generating natural language description;

[0006] Step S2: using BGE-large-zh word embedding model to encode the natural language description into a low-dimensional semantic vector;

[0007] Step S3: constructing a semantic keyword association graph based on Node2Vec graph embedding technology, and constructing a special semantic tree for geographic features by fusing general semantic knowledge base and geographic field ontology;

[0008] Step S4: screening a candidate subset according to the preset geographic feature priority and query image saliency detection result;

[0009] Step S5: extracting global features and local key point descriptors of the images in the candidate subset;

[0010] Step S6: constructing a visual similarity calculation model based on Mamba state space model, and modeling feature sequence dependency relationship;

[0011] Step S7: fusing semantic matching score and visual similarity, using weighted ranking strategy to output the final matching result, and cooperatively accelerating according to the matching result combined with the lightweight E-Swin Transformer model.

[0012] In some embodiments, in step S1, the instruction fine-tuning comprises:

[0013] Using field customized instructions to explicitly constrain semantic generation direction, and generating natural language description containing road type, building feature and landmark.

[0014] In some embodiments, in step S3, the semantic keyword association graph constructed based on Node2Vec graph embedding technology comprises:

[0015] Establishing semantic association and introducing position saliency weight, and dynamically adjusting retrieval priority.

[0016] In some embodiments, in step S4, the screening of the candidate subset comprises:

[0017] Compressing the candidate set based on Top-K screening strategy;

[0018] Triggering secondary retrieval process for low confidence candidates, and optimizing candidate set by expanding synonyms or relaxing threshold.

[0019] In some embodiments, step S5 includes extracting global features and local keypoint descriptors of the images in the candidate subset using a lightweight E-Swin Transformer model.

[0020] In some embodiments, step S5 further includes:

[0021] replacing traditional MLP with depthwise separable convolution;

[0022] replacing dot product with linear transformation in self-attention layer to reduce feature extraction time.

[0023] In some embodiments, in step S6, the visual similarity calculation model includes:

[0024] combining local window attention with cross-window connection to balance local details and global semantic feature discriminability, and using linear complexity to model feature sequence dependency.

[0025] In some embodiments, the feature sequence dependency is determined according to the following formula:

[0026] Sim(f q ,f r )=Softmax(Mamba(f q ,f r ))

[0027] wherein Mamba(f q ,f r ) represents a state space model.

[0028] In some embodiments, in step S7, the weighted ranking strategy includes:

[0029] initially setting a semantic matching score weight, and dynamically adjusting the weight through online learning of the edge device;

[0030] using Faiss HNSW index to perform approximate nearest neighbor search to reduce time complexity.

[0031] In some embodiments, the weighted ranking strategy is determined according to the following formula:

[0032] S final =λ·S s +(1-λ)·S v

[0033] wherein λ is a weight, S S is a fused semantic matching score, and S v is a visual similarity.

[0034] The beneficial effects of the above technical solution are as follows:

[0035] This invention achieves significant improvements in multiple aspects through a hierarchical architecture design of "semantic pre-screening - visual fine matching" and optimization of key technical features. In terms of computational efficiency, it employs BGE-large-zh to encode image semantics into a 256-dimensional vector, reducing the dimension by more than 90% compared to traditional pixel-level features. Dynamic weighting and a Top-K strategy compress candidate images by more than 80%. The improved E-Swin Transformer has a parameter count controlled within 21M, reducing FLOPs by 35%. Combined with the linear complexity of the Mamba model, the matching speed is twice that of the traditional Transformer. Finally, the processing time per frame is controlled within 150ms, more than three times faster than traditional methods, approaching the real-time threshold for scenarios such as autonomous driving, thus solving the problem of computational bottlenecks restricting practical application. In terms of robustness in cross-view positioning, a high-level description is generated based on BLIP-2, and accurate semantic alignment of ground features across views is achieved by combining geographic information semantic trees and Node2Vec graph embedding. Secondary retrieval accommodates feature loss caused by dynamic occlusion, and the fusion of global and local features enhances generalization ability in unstructured environments. Injecting spatial attributes into the geographic semantic tree improves matching accuracy in complex scenes, effectively addressing differences between sky and ground views, lighting variations, and dynamic occlusion. Regarding lightweight design and generalization, the semantic description-based matching logic reduces reliance on pixel-level details, saving over 70% in storage and transmission costs. The overall architecture parameters are controlled within 50M, allowing operation on embedded devices without dedicated computing chips. The online learning mechanism for dynamic weights reduces adaptation costs by over 60%, providing an efficient and reliable solution for cross-view geolocation in GNSS-free environments. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1 A flowchart illustrating a cross-view geolocation acceleration method based on semantic description, provided as an embodiment of the present invention;

[0038] Figure 2 A lightweight E-Swin Transformer model architecture diagram of a cross-view geolocation acceleration method based on semantic description is provided as an embodiment of the present invention;

[0039] Figure 3The following is a framework diagram of the Mamba state space model for a cross-view geolocation acceleration method based on semantic description, provided as an embodiment of the present invention. Detailed Implementation

[0040] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0041] Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar symbols denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the invention, and should not be construed as limiting the invention.

[0042] Example 1

[0043] One embodiment of the present invention provides a cross-view geolocation acceleration method based on semantic description, referring to... Figures 1-3 As shown, it includes:

[0044] Step S1: Fine-tune the input query image and database image using the BLIP-2 visual language model to generate a natural language description;

[0045] In a specific embodiment of the present invention, step S1, the instruction fine-tuning includes:

[0046] By using domain-specific custom instructions to explicitly constrain the semantic generation direction, natural language descriptions containing road types, building features, and landmarks are generated.

[0047] Step S2: Encode the natural language description into a low-dimensional semantic vector using the BGE-large-zh word embedding model;

[0048] Specifically, the BLIP-2 visual language model is used to fine-tune the input query images (such as real-time images from vehicle cameras) and database images (such as satellite imagery) using instructions (Prompt: "Generate a description including road type, building features, and landmarks") to generate natural language descriptions (e.g., "two-way six-lane road, glass-walled high-rise building, cross-river bridge"). Domain-customized instructions (Prompt) explicitly constrain the direction of semantic generation, improving the completeness of geographic entity descriptions. Simultaneously, the BGE-large-zh word embedding model encodes the natural language descriptions into 256-dimensional low-dimensional semantic vectors, preserving core semantic information such as "road-building-landmark," reducing the dimensionality of subsequent retrieval (by more than 90% compared to traditional pixel-level feature dimensions).

[0049] Based on Node2Vec graph embedding technology, a semantic keyword association graph is constructed, such as establishing semantic associations for "building-landmark" and "road-lane" to support fuzzy matching. This allows the retrieval of images related to "Twin Towers" when searching for "Twin Towers," solving the terminology omission problem of traditional exact matching. At the same time, a general semantic knowledge base and a geographic domain ontology (OpenStreetMap tag system) are integrated to construct a feature-specific semantic tree. For example, under the "bridge" node, subclasses such as "cable-stayed bridge" and "suspension bridge" are expanded, and spatial relationship attributes such as "adjacent" and "containment" are injected to give it the ability to subdivide geographic entities and express spatial logic, enhancing the granularity and robustness of semantic retrieval. In addition, location salience weights are introduced, such as increasing the weight of keywords related to building areas by 50%. The retrieval priority is dynamically adjusted according to the contribution of geographic entities to the location, avoiding low-value semantics such as "vegetation" from dominating the retrieval process and effectively improving the quality of the candidate set.

[0050] Step S3: Construct a semantic keyword association graph based on Node2Vec graph embedding technology, and integrate a general semantic knowledge base with a geographic domain ontology to construct a semantic tree specifically for geographic features;

[0051] In a specific embodiment of the present invention, step S3, which involves constructing a semantic keyword association graph based on Node2Vec graph embedding technology, includes:

[0052] Establish semantic associations between "buildings-landmarks" and "roads-lanes" and introduce location salience weights to dynamically adjust search priorities.

[0053] Step S4: Select a candidate subset based on the preset land cover priority and the saliency detection results of the query image;

[0054] In a specific embodiment of the present invention, step S4, the screening of the candidate subset includes:

[0055] Compress the candidate set based on the Top-K screening strategy;

[0056] For candidates with low confidence that trigger a secondary search process, the candidate set is optimized by expanding synonyms or relaxing the threshold.

[0057] Specifically, based on the predefined feature priority (OpenStreetMap tag system) and the saliency detection results of the query image, the comprehensive score of keywords is calculated, and a candidate subset containing ≥2 high-weight keywords is selected. Then, the candidate set is compressed by more than 80% through the Top-K screening strategy (K=500) to reduce the amount of subsequent visual matching calculation. At the same time, for low-confidence candidates (matching score <0.7), a secondary search process is triggered, which is processed by expanding synonyms (such as "bridge" → "cable-stayed bridge") or relaxing the threshold (reducing the number of matching keywords from 2 to 1), thereby improving the accuracy and comprehensiveness of candidate set screening and optimizing the preliminary data processing process of cross-view geolocation.

[0058] Protection points: Based on the dynamic weight allocation strategy of ground feature priority and saliency detection, combined with the candidate set optimization method of secondary retrieval, efficient semantic pre-screening is achieved.

[0059] Step S5: Extract global features and local key point descriptors from the images in the candidate subset;

[0060] In a specific embodiment of the present invention, step S5 includes: using a lightweight E-Swin Transformer model to extract global features and local key point descriptors of the images in the candidate subset.

[0061] In some specific embodiments, step S5 further includes:

[0062] Replace traditional MLP with depthwise separable convolution;

[0063] In the self-attention layer, linear transformations are used instead of dot products to calculate weights, in order to reduce the time required for feature extraction.

[0064] Specifically, for candidate subset C cand Each satellite image in I r and query image I q An improved E-SwinTransformer module is used for feature extraction: depthwise separable convolutions replace traditional MLPs, reducing computational cost (FLOPs) by 35% and keeping the number of parameters below 21M; linear transformations are used instead of dot products to calculate weights in the self-attention layer, avoiding weight decay issues during long sequence matching, reducing feature extraction time by 40% compared to traditional Transformers. The output includes global features f. q (512-dimensional) and local keypoint descriptor f r (Based on the Super-Point algorithm, an average of 500 key points are extracted from a single image).

[0065] Step S6: Construct a visual similarity calculation model based on the Mamba state space model to model the dependency relationship of feature sequences;

[0066] Specifically, a visual similarity calculation module is constructed based on the Mamba State-Space Model (SSM) to model feature sequence dependencies with linear complexity O(N), as shown in the following formula:

[0067] Sim(f q ,f r =Softmax(Mamba(f q ,f r ))

[0068] By combining local window attention (W-MHSA) with cross-window connectivity, the feature discriminativeness of local details and global semantics is balanced, and the matching speed is improved by 2 times compared with the traditional Transformer.

[0069] Fusion semantic matching score S s Visual similarity S v A weighted sorting strategy is adopted:

[0070] S final =λ·S s +(1-λ)·S v

[0071] The initial weight λ is set to 0.6 (semantic proportion), which can be dynamically adjusted through online learning on edge devices. The Faiss HNSW index is used to perform an approximate nearest neighbor search on the candidate subset, reducing the time complexity from O(N) to O(N). 2 The time complexity is reduced to O(NlogN), and the single-frame matching time is controlled within 150ms.

[0072] Protection point: A lightweight E-Swin Transformer + Mamba model is adopted to balance computational efficiency and feature discriminativeness;

[0073] Step S7: Integrate semantic matching score and visual similarity, and output the final matching result using a weighted ranking strategy. Based on the matching result, combine the lightweight E-Swin Transformer model for collaborative acceleration.

[0074] In a specific embodiment of the present invention, step S6, the visual similarity calculation model includes:

[0075] By combining local window attention and cross-window connectivity, the discriminative power of features is balanced between local details and global semantics, and linear complexity is used to model feature sequence dependencies.

[0076] In a specific embodiment of the present invention, the feature sequence dependency is determined according to the following formula:

[0077] Sim(f q ,f r=Softmax(Mamba(f q ,f r ))

[0078] In the formula, Mamba(f q ,f r ) represents the state-space model.

[0079] In a specific embodiment of the present invention, step S7 includes the following weighted sorting strategy:

[0080] Initially set semantic matching score weights, and dynamically adjust the weights through online learning on edge devices;

[0081] By using the Faiss HNSW index for approximate nearest neighbor search, the time complexity is reduced.

[0082] In one specific embodiment of the present invention, the weighted sorting strategy is determined according to the following formula:

[0083] S final =λ·S s +(1-λ)·S v

[0084] In the formula, λ is the weight, and S S To integrate semantic matching scores, S v Visual similarity.

[0085] Compared with existing technologies, this invention achieves significant improvements in multiple aspects through a hierarchical architecture design of "semantic pre-screening - visual fine matching" and optimization of key technical features. In terms of computational efficiency, it uses BGE-large-zh to encode image semantics into a 256-dimensional vector, reducing the dimension by more than 90% compared to traditional pixel-level feature dimensions. Through dynamic weight screening and Top-K strategy, candidate images are compressed by more than 80%. The improved E-Swin Transformer has a parameter size controlled within 21M, reducing FLOPs by 35%. Combined with the linear complexity of the Mamba model, the matching speed is twice that of the traditional Transformer. Finally, the processing time per frame is controlled within 150ms, which is more than three times faster than traditional methods, approaching the real-time threshold of scenarios such as autonomous driving, thus solving the problem of computational bottlenecks restricting practical application. In terms of robustness in cross-view positioning, a high-level description is generated based on BLIP-2, and accurate semantic alignment of ground features across views is achieved by combining geographic information semantic trees and Node2Vec graph embedding. Secondary retrieval accommodates feature loss caused by dynamic occlusion, and the fusion of global and local features enhances generalization ability in unstructured environments. Injecting spatial attributes into the geographic semantic tree improves matching accuracy in complex scenes, effectively addressing differences between sky and ground views, lighting variations, and dynamic occlusion. Regarding lightweight design and generalization, the semantic description-based matching logic reduces reliance on pixel-level details, saving over 70% in storage and transmission costs. The overall architecture parameters are controlled within 50M, allowing operation on embedded devices without dedicated computing chips. The online learning mechanism for dynamic weights reduces adaptation costs by over 60%, providing an efficient and reliable solution for cross-view geolocation in GNSS-free environments.

[0086] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

[0087] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. The embodiments of the present invention are described with reference to flowchart illustrations and / or block diagrams of methods, terminal devices (systems), and computer program products according to embodiments of the present invention. It should 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 terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 The computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate 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 1 One or more processes and / or boxes Figure 1 The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing terminal equipment to cause a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1The steps of the functions specified in one or more boxes. Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the embodiments of the invention. Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.

[0088] The methods and apparatus provided by the present invention have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

[0089] In the description of this specification, references to terms such as "an embodiment," "some embodiments," "example," "specific example," or "a specific embodiment" or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0090] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A cross-view geolocation acceleration method based on semantic description, characterized in that, include: Step S1: Fine-tune the input query image and database image using the BLIP-2 visual language model to generate a natural language description; Step S2: Encode the natural language description into a low-dimensional semantic vector using the BGE-large-zh word embedding model; Step S3: Construct a semantic keyword association graph based on Node2Vec graph embedding technology, and integrate a general semantic knowledge base with a geographic domain ontology to construct a semantic tree specifically for geographic features; Step S4: Select a candidate subset based on the preset land cover priority and the saliency detection results of the query image; Step S5: Extract global features and local key point descriptors from the images in the candidate subset; Step S6: Construct a visual similarity calculation model based on the state-space model to model the dependency relationship of feature sequences; Step S7: Integrate semantic matching score and visual similarity, and output the final matching result using a weighted ranking strategy. Accelerate the process by combining the matching result with the lightweight E-Swin Transformer model.

2. The cross-view geolocation acceleration method based on semantic description according to claim 1, characterized in that, In step S1, the instruction fine-tuning includes: By using domain-specific custom instructions to explicitly constrain the semantic generation direction, natural language descriptions containing road types, building features, and landmarks are generated.

3. The cross-view geolocation acceleration method based on semantic description according to claim 1, characterized in that, In step S3, the construction of the semantic keyword association graph based on Node2Vec graph embedding technology includes: Establish semantic associations and introduce positional salience weights to dynamically adjust retrieval priorities.

4. The cross-view geolocation acceleration method based on semantic description according to claim 1, characterized in that, In step S4, the candidate subset includes: Compress the candidate set based on the Top-K screening strategy; For candidates with low confidence that trigger a secondary search process, the candidate set is optimized by expanding synonyms or relaxing the threshold.

5. The cross-view geolocation acceleration method based on semantic description according to claim 1, characterized in that, Step S5 includes: using a lightweight E-Swin Transformer model to extract global features and local keypoint descriptors of the images in the candidate subset.

6. The cross-view geolocation acceleration method based on semantic description according to claim 5, characterized in that, Step S5 further includes: Replace traditional MLP with depthwise separable convolution; In the self-attention layer, linear transformations are used instead of dot products to calculate weights, in order to reduce the time required for feature extraction.

7. The cross-view geolocation acceleration method based on semantic description according to claim 1, characterized in that, In step S6, constructing the visual similarity calculation model includes: By combining local window attention and cross-window connectivity, the discriminative power of features is balanced between local details and global semantics, and linear complexity is used to model feature sequence dependencies.

8. The cross-view geolocation acceleration method based on semantic description according to claim 7, characterized in that, The feature sequence dependency is determined according to the following formula: Sim(f q ,f r )=Softmax(Mamba(f q ,f r )) In the formula, Mamba(f q ,f r ) represents the state-space model.

9. The cross-view geolocation acceleration method based on semantic description according to claim 1, characterized in that, In step S7, the weighted sorting strategy includes: Initially set semantic matching score weights, and dynamically adjust the weights through online learning on edge devices; By using the Faiss HNSW index for approximate nearest neighbor search, the time complexity is reduced.

10. The cross-view geolocation acceleration method based on semantic description according to claim 1, characterized in that, The weighted sorting strategy is determined according to the following formula: S final =λ·S s +(1-λ)·S v In the formula, λ is the weight, and S S To integrate semantic matching scores, S v Visual similarity.

Images