A text-based semantic-guided cross-view geolocation system and method for low-light UAV images obscured by clouds.

Abstract

This invention relates to the field of UAV positioning technology, specifically to a text-based semantic-guided cross-view geolocation system and method for UAV images in low-light conditions caused by cloud cover. The system includes a UAV low-light image input module for acquiring real-time images collected by the UAV under cloud cover conditions; and a cloud cover and low-light degree perception module for calculating normalized brightness distribution, local contrast, dark area ratio, edge response, and fogging indices on the real-time images, and calculating a cloud cover low-light score α based on these indices. This invention adapts to complex environmental changes by decomposing occlusion, image distortion, radiation differences, and seasonal variations into perceptible, augmentable, and semantically correctable sub-problems. The system can output more stable satellite candidate positions in scenarios such as low-altitude inspection, emergency rescue, disaster assessment, and autonomous navigation.

Description

Technical Field

[0001] This invention relates to the field of drone positioning technology, specifically to a text-based semantic-guided cross-view geolocation system and method for drone images in low-light conditions obscured by clouds. Background Technology

[0002] UAV geolocation typically involves searching and matching real-time low-altitude UAV images with a pre-built satellite image database. When GPS signals are blocked, spoofed, interfered with, or unavailable, this type of visual positioning technology can provide location estimation for urban inspections, emergency rescue, low-altitude aerial surveying, autonomous navigation, and disaster situational awareness. Existing cross-view positioning methods mostly employ CNNs, Transformers, or visual-language pre-trained models to extract features from UAV and satellite images, and then perform similarity retrieval through metric learning or contrastive learning.

[0003] Existing technologies can be broadly categorized into four types: First, low-light image enhancement and scene matching schemes alleviate matching difficulties in low light conditions through image enhancement, local feature extraction, feature matching, and homography matrix estimation; second, severe weather cross-view localization networks improve feature representation under conditions such as rain, fog, and snow through weather simulation, directional attention, and multi-branch feature fusion; third, semantic-driven cross-view matching schemes use CLIP image-text similarity to generate semantic triggers or semantic masks, and then constrain local feature matching regions; fourth, visual language model cross-view geolocation methods utilize CLIP to generate text descriptions and improve the consistency between UAV views and satellite views through multimodal fusion and various loss functions.

[0004] The above-mentioned schemes demonstrate that low-light enhancement, robust features in adverse weather, semantic priors, and multimodal losses are all relevant to this field, but they still have shortcomings: low-light enhancement schemes often treat enhancement as a preprocessing step and lack explicit quantification of cloud cover levels; adverse weather simulation schemes typically use rain, fog, and snow as training perturbations and do not establish interpretable scores for real-time low-light degradation caused by cloud cover; semantic masking schemes are more biased towards local target matching and do not use illumination-invariant text semantics as a signal for repairing UAV dark area features; visual-language cross-view localization schemes emphasize text description guidance but do not unify low-light enhancement intensity, semantic correction weights, and cross-view retrieval loss into the same computational closed loop.

[0005] Therefore, a cross-view geolocation technology for cloud-covered, low-light UAV images is needed: it should be able to detect and quantify low-light degradation, recover structural cues that contribute to matching without excessively amplifying noise, and correct degraded visual features by leveraging the land cover attributes in the text semantics that are insensitive to differences in illumination, season, and radiation. To this end, we propose a text semantic-guided cross-view geolocation system and method for cloud-covered, low-light UAV images to solve the above problems. Summary of the Invention

[0006] This invention aims to solve the problem of stable matching between low-altitude tilted UAV images and overhead satellite images in cross-view geolocation of UAVs. Common challenges in this field include: geometric projection differences and image distortion caused by low-altitude tilted perspectives; structural misalignment caused by different flight altitudes, orientations, scales, and resolutions; local information loss caused by obstructions such as buildings, tree canopies, and clouds; radiation differences caused by different sensors, imaging times, and atmospheric conditions; and changes in vegetation, shadows, and surface appearance due to seasonal variations.

[0007] In addressing the aforementioned challenges, this invention further focuses on more specific and frequent adverse weather scenarios: when drones encounter cloud cover, fog, or rain during low-altitude inspections, emergency rescues, or disaster assessments, real-time acquired images exhibit low brightness, reduced contrast, blurred dark textures, and diminished details in road markings and building facades—a form of dark light degradation. This degradation significantly attenuates the local gradients and high-frequency texture features extracted by visible light visual encoders such as CLIP ViT-B, making it difficult to align road skeletons, building boundaries, vegetation outlines, and water edges—which would otherwise be suitable for cross-view matching—with satellite images.

[0008] The purpose of this invention is to provide a method and system that can sense the degree of cloud cover and low light degradation, adaptively enhance UAV low light images, and use the illumination invariant attribute in satellite text semantics to repair the visual features of UAVs. This improves the reliability of cross-view geolocation from UAVs to satellite image databases and the clarity of authorized protection boundaries when occlusion, distortion, radiation differences and seasonal changes coexist.

[0009] To achieve the above objectives, the present invention provides the following technical solution: a text semantic-guided cross-view geolocation system for low-light UAV images obscured by clouds, comprising:

[0010] The drone low-light image input module is used to acquire real-time images collected by the drone under cloud cover conditions.

[0011] The cloud cover and dim light level perception module is used to calculate the normalized brightness distribution, local contrast, dark area ratio, edge response and fogging index of the real-time image, and calculate the cloud cover dim light score α based on the above indexes.

[0012] An adaptive low-light enhancement module is used to generate a spatially varying low-light enhancement mask based on the cloud cover low-light score α, and to perform edge-preserving adaptive enhancement of the UAV image.

[0013] The triplet text description generation module is used to generate triplet text descriptions for satellite images offline.

[0014] The image encoding module uses a pre-trained CLIP ViT-B image encoder to extract visual features from the enhanced UAV images and satellite images;

[0015] The text encoding module uses a pre-trained BERT text encoder to extract semantic features from satellite text descriptions;

[0016] The semantically guided feature repair module is used to generate semantically repaired UAV features based on the cloud cover dark light score α, the UAV visual features before enhancement, the UAV visual features after enhancement, satellite text semantic features, and satellite visual features, through a gating mechanism.

[0017] The bidirectional image-to-text / image-to-image alignment module is used to perform bidirectional alignment of UAV images to satellite images, satellite images to UAV images, UAV images to satellite text, and satellite text to UAV images.

[0018] The feature fusion retrieval module is used to fuse the repaired UAV features with satellite text features to generate a UAV query fusion vector, fuse satellite visual features with satellite text features to generate a satellite candidate fusion vector, and calculate the similarity.

[0019] The contrastive learning training module is used to train the system end-to-end using bidirectional contrastive learning loss.

[0020] Based on a text-semantic guided cross-view geolocation system for cloud-obscured, low-light UAV images, this invention also discloses a text-semantic guided cross-view geolocation method for cloud-obscured, low-light UAV images, comprising the following steps:

[0021] Step 1: Acquire real-time images taken by the drone under cloud cover conditions. and a pre-built georeferenced satellite candidate image library, wherein each satellite image in the satellite candidate image library All are associated with pre-generated offline triple text descriptions. ;

[0022] For the real-time image Calculate its normalized brightness distribution Local contrast normalization value Dark area ratio Normalized edge response and atomization index The cloud cover dark light score is calculated using the following formula. :

[0023]

[0024]

[0025] in, A collection of image pixels. Let σ(·) be the threshold for the dark area, and let σ(·) be the Sigmoid function. , , , , and These are learnable parameters or validated weights;

[0026] Step 2: Based on the cloud cover dark light score α and the local brightness map, generate the dark light enhancement mask Mu(x) according to the following formula and perform adaptive enhancement on the UAV image:

[0027]

[0028]

[0029] in, G(·) is the edge protection coefficient, G(·) is the edge preservation enhancement function, and clip(·, 0, 1) is the cutoff function.

[0030] Step 3: Extract the enhanced UAV images using the pre-trained CLIP ViT-B image encoder φI(·). visual features and satellite images visual features The pre-trained BERT text encoder φT(·) is used to extract triplet text descriptions of satellite images. semantic features :

[0031]

[0032] This formula maps the UAV visual information, satellite visual information, and illumination-invariant text semantics before and after low-light enhancement to a unified feature space, providing input for subsequent semantic repair and cross-view retrieval;

[0033] The image encoder and text encoder remain frozen during training;

[0034] Step 4: Based on the cloud cover low light score α and the enhanced visual features of the UAV... Enhanced visual features of drones Satellite text semantic features and satellite visual features Generate semantically repaired drone features using the following formula. :

[0035]

[0036]

[0037] Where [·;·] represents feature concatenation, and Here are the parameters of the gated network, and ψ(·) represents the semantic correction network. This represents element-wise multiplication, and Norm(·) is the normalization function.

[0038] Step 5: Extract the semantically repaired drone features With satellite text semantic features The concatenated vectors are input into a multilayer perceptron to generate a UAV query fusion vector. Satellite visual features With satellite text semantic features The concatenated data is then input into a multilayer perceptron to generate satellite candidate fusion vectors. ;

[0039] The similarity between the UAV query fusion vector and the candidate satellite fusion vectors is calculated using the following formula.

[0040]

[0041] in, For drone query fusion vectors, For satellite candidate fusion vectors;

[0042] Step Six: Based on similarity The satellite candidate image library is sorted from highest to lowest, and the top-K satellite candidate images and their corresponding geographic coordinates are output.

[0043] More preferably, the triplet text description The generation method is as follows: For each satellite candidate image, the multimodal large language model Qwen2-VL is called offline to first generate attribute prompts, and then the attribute prompts are input into the multimodal large language model to generate fine-grained text descriptions; the attribute prompts include building layout, road layout and building style; the fine-grained text descriptions include building shape, road direction, spatial relationship and building appearance; the generated prompts explicitly require priority to describe attributes that are constant in illumination and relatively stable in season, including road framework, building layout, roof outline, water body boundary, playground shape, open area and vegetation edge.

[0044] More preferably, the edge preservation enhancement function G(·) is implemented using guided filtering or bilateral filtering.

[0045] More preferably, the semantic correction network ψ(·) is implemented using a cross-attention mechanism to enhance the post-UAV features. As a query, using satellite text semantic features and satellite visual features The concatenation of values ​​serves as the key and value, and semantically enhanced drone features are obtained through multi-head cross-attention computation.

[0046] Furthermore, the following loss function is used for optimization during the training phase:

[0047]

[0048]

[0049]

[0050]

[0051] in, For the two-way alignment loss between UAV imagery and satellite imagery, For the semantic alignment loss of UAV image-satellite text, To repair the consistency constraint loss for features, Classify losses for optional locations. For temperature coefficient, , , , Let P(·) be the loss weight and P(·) be the projection function.

[0052] Further optimized, during the training phase, cloud-masked dark light enhancement samples are constructed by reducing brightness, local occlusion, fogging perturbation, and contrast reduction.

[0053] Compared with the prior art, the beneficial effects of the present invention are:

[0054] 1. The technical problem addressed by this invention is more specific. Compared with methods that only generally address cross-viewpoint differences, this invention focuses on protecting against the degradation of UAV images in low light caused by cloud cover, fog, or rain, and clarifies the specific technical problems of low brightness, decreased contrast, blurred dark textures, and degradation of visual encoder features.

[0055] 2. The technical solution of this invention has a computable closed loop. The occlusion dark light score α is calculated by the dark area ratio, local contrast, edge response and fogging index. Then, α controls the enhancement mask Mu(x), semantic repair g and consistency loss Lcon, so that each link of "perception-enhancement-repair-retrieval-training" has a clear mathematical relationship.

[0056] 3. This invention provides targeted technical means. The system does not simply rely on an ideal model library or general image enhancement. Instead, it forms a complete closed loop through dark light occlusion scoring, adaptive enhancement, residual fidelity, text semantic feature repair, and bidirectional alignment, which can adopt different enhancement and correction strategies for different degrees of occlusion.

[0057] 4. This invention reduces noise and artifact interference. General histogram equalization or Retinex may over-enhance noise in dark areas. This invention utilizes semantic prior constraints such as road skeletons, building layouts, water body boundaries, and vegetation edges to enhance and repair directions, making the restored features more suitable for cross-view matching.

[0058] 5. This invention adapts to complex environmental changes. By breaking down occlusion, image distortion, radiation differences, and seasonal changes into perceptible, augmentable, and semantically correctable sub-problems, the system can output more stable satellite candidate positions in scenarios such as low-altitude inspection, emergency rescue, disaster assessment, and autonomous navigation.

[0059] 6. This invention is easy to deploy in engineering. Satellite image features and triplet text semantics can be generated and cached offline. The UAV only needs to perform low-light perception, enhancement, feature extraction and retrieval sorting in real time. It is suitable for deployment on airborne edge computing devices or ground station systems. Attached Figure Description

[0060] Figure 1 This is the overall network topology diagram of the VLGeo system for low-light scenarios under cloud cover, illustrating the connections between UAV low-light images, cloud cover and low-light perception, adaptive enhancement, CLIP ViT-B image encoding, Qwen2-VL triple text generation, BERT text encoding, semantically guided feature repair, bidirectional alignment, MLP fusion, and Top-K geolocation output.

[0061] Figure 2This diagram illustrates the cloud-occluded low-light enhancement and feature repair module, showing how inputs with low brightness, reduced contrast, and blurred dark textures are processed through occlusion scoring, local brightness maps, edge preservation enhancement, noise suppression constraints, and text semantic correction to form repaired drone features.

[0062] Figure 3 This diagram illustrates the generation of textual semantics from satellite image triples and the extraction of illumination-invariant attributes. It demonstrates how Qwen2-VL generates triples of target features, visibility attributes, and spatial relationships from satellite images and encodes them into textual semantic features that can be used for cross-view correction using BERT.

[0063] Figure 4 This is a flowchart of the training and inference process of the present invention, showing the complete process of constructing dark light samples, bidirectional InfoNCE contrastive learning of images and text, feature repair consistency constraints, and the inference process from real-time UAV images to Top-K satellite candidate positions.

[0064] Figure 5 This is a schematic diagram of a multimodal feature fusion and semantic gating network robust to low light conditions.

[0065] Figure 6 This is a schematic diagram of Top-K retrieval and candidate validation for low-light drone images. Detailed Implementation

[0066] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0067] Example

[0068] Please see Figures 1 to 6 This invention provides a text-based semantic-guided cross-view geolocation system and method for low-light drone images obscured by clouds. The system includes: a drone low-light image input module, a cloud obscuration and low-light degree perception module, an adaptive low-light enhancement module, a triplet text description generation module, an image encoding module, a text encoding module, a semantic-guided feature repair module, a bidirectional image-text / image-image alignment module, a feature fusion retrieval module, and a contrastive learning training module.

[0069] The system uses real-time UAV images as query input and a satellite candidate image library and its text descriptions generated by a large visual language model as the retrieval base. To address low brightness, low contrast, and texture degradation caused by cloud cover, the system first estimates the dark light degradation weights of the UAV images and then performs edge-preserving adaptive enhancement. Subsequently, it extracts enhanced UAV and satellite visual features using CLIP ViT-B, extracts satellite textual semantic features using BERT, and corrects the UAV features using stable ground features, spatial relationships, and illumination-invariant properties from the textual semantics. Finally, it generates cross-view matching vectors through bidirectional alignment and MLP fusion, outputting the Top-K satellite candidate positions.

[0070] II. Key Symbols and Formulas

[0071] To ensure that the technical solution forms an implementable computational closed loop, the present invention defines the following symbols: For real-time images of drones, For satellite candidate images, This is the satellite triplet text generated by Qwen2-VL; The normalized brightness at pixel x. A collection of image pixels. This is the threshold for the dark area; This is the local contrast normalization value. For the dark area ratio, To normalize the edge response, This refers to indicators for atomized or low-saturation regions. Rate the darkness caused by cloud cover; φI(·) is the low-light enhancement mask; G(·) is the edge-preserving enhancement function; φI(·) is the CLIP ViT-B image encoder, and φT(·) is the BERT text encoder; , , , , These represent the original UAV features, enhanced UAV features, satellite visual features, satellite text features, and semantically repaired UAV features, respectively; σ(·) is the Sigmoid function, Norm(·) is the normalization function, and ⊙ represents element-wise multiplication.

[0072] (1) Dark area ratio and cloud cover low light score:

[0073]

[0074]

[0075] in, , , , , and These are learnable parameters or validated weights. This formula unifies the reduction in brightness, contrast, dark area expansion, edge response decay, and low saturation due to clouds and fog into a dark light rating of 0 to 1. The larger the value, the more the drone image needs enhancement and semantic repair.

[0076] (2) Adaptive low-light enhancement mask and enhanced image:

[0077]

[0078]

[0079] in, This is the edge protection coefficient. This formula ensures that enhancement primarily occurs in low-brightness areas with weak structural response; when image brightness is normal or edges are clear... When the value is close to 0, the system retains the original image, avoiding overexposure, noise amplification, or artifacts caused by general enhancement.

[0080] (3) Image and text feature encoding:

[0081]

[0082] This formula maps the UAV visual information, satellite visual information, and illumination-invariant text semantics before and after low-light enhancement to a unified feature space, providing input for subsequent semantic repair and cross-view retrieval.

[0083] (4) Semantic-guided feature repair gating:

[0084]

[0085]

[0086] Where [·;·] represents feature concatenation, Wg and bg are gating parameters, and ψ(·) is the semantic correction network. At higher levels, The system increases its use of light-invariant textual semantics, such as road frameworks, building layouts, water body boundaries, and vegetation edges, to compensate for missing clues in dark areas. At lower levels, By reducing the size of the image, the system retains more of the enhanced visual features.

[0087] (5) Integrating retrieval vectors and similarity:

[0088]

[0089]

[0090] in, For drone query fusion vectors, These are satellite candidate fusion vectors. The inference phase follows... Sort the satellite images from highest to lowest and output the Top-K candidate images and their geographic coordinates.

[0091] (6) Training loss function:

[0092]

[0093]

[0094]

[0095]

[0096] in, For the two-way alignment loss between UAV imagery and satellite imagery, For the semantic alignment loss of UAV image-satellite text, For feature repair consistency constraints, The classification loss is for the selectable location, where τ is the temperature coefficient. to The total loss is used as the loss weight. This total loss enables the model to simultaneously learn cross-view visual consistency, textual semantic consistency, and dark lighting restoration stability.

[0097] III. Cloud cover and perception of low light levels

[0098] For the input UAV image Iu, the system calculates brightness distribution, local contrast, proportion of dark areas, edge response intensity, proportion of fogged / low-saturation regions, and feature entropy to obtain a dark light occlusion score α. α represents the degree to which the image is affected by cloud cover, fog, or rain. A high α indicates the presence of large dark areas or texture degradation in the image, requiring an enhancement module to improve structural visibility and a semantic insulation module to increase the correction weight of textual semantics on visual features. A low α indicates that the system retains more original visual features, avoiding unnecessary enhancements that introduce noise and artifacts.

[0099] IV. Adaptive Dark Light Enhancement

[0100] Based on the low-light occlusion score α and the local brightness map, the system performs adaptive enhancement on the UAV image. Instead of using fixed histogram stretching, the enhancement process differentiates between low-brightness, edge, and high-noise regions: contrast restoration is performed on structural areas such as roads, roof outlines, building edges, and vegetation boundaries; noise suppression constraints are applied to large areas of textureless dark areas and cloud / fog noise areas; and upper limits are set for overly bright areas to avoid artifacts. This processing outputs an enhanced image Iu', while retaining original image features as residual branches to suppress over-enhancement risks during subsequent feature repair.

[0101] V. Generation of Triple Text Descriptions

[0102] Offline, Qwen2-VL is used to generate triplet text descriptions for satellite candidate images. Each triplet includes: target feature or region category, visibility attribute, and spatial relationship. Examples include "road-surrounding-building complex", "playground-adjacent-woodland", and "water body-located-east of road". The generation prompts explicitly require prioritizing descriptions of attributes with constant illumination and relatively stable seasons, including road framework, building layout, roof outline, water body boundary, playground shape, open areas, and vegetation edges, rather than relying on transient shadows, brightness, or color.

[0103] VI. Image and Text Feature Encoding

[0104] CLIP ViT-B was used to extract the visual features fue of the enhanced UAV image Iu' and the visual features fs of the satellite image Is; BERT was used to extract the semantic features ft of the triplet text description. The satellite image features and text features can be pre-cached offline, and the UAV features are computed in real time. The original image features and the enhanced image features are preserved for the UAV image, and gated fusion is performed based on the dark light occlusion score α.

[0105] VII. Semantic Guidance Feature Repair

[0106] The semantic-guided feature restoration module generates restoration features *fr* based on the low-light occlusion score α, UAV visual features, satellite textual semantic features, and satellite visual features. When α is high, the module increases the correction weight of textual semantics for UAV features, and uses illumination-invariant properties such as road direction, relative location of building clusters, shape of open areas, and water / vegetation boundaries to compensate for missing discriminative information in dark areas. When α is low, the module mainly relies on visual features, with textual semantics serving only as an auxiliary constraint. This avoids the amplification of noise in dark areas by general enhancement methods while maintaining semantic consistency with satellite imagery.

[0107] VIII. Bidirectional Semantic Alignment and Feature Fusion

[0108] The system simultaneously performs bidirectional alignment: from UAV image to satellite image, from satellite image to UAV image, from UAV image to satellite text, and from satellite text to UAV image. A visual branch constrains geometric structure and texture correspondence, while a text branch constrains illumination-invariant semantics and spatial relationships. Subsequently, the repaired UAV features, satellite visual features, and satellite text features are fused using MLP to obtain a unified matching vector, and the satellite candidate library is ranked using cosine similarity or equivalent similarity functions.

[0109] IX. Training and Reasoning Process

[0110] During the training phase, the triplet of UAV image, corresponding satellite image, and satellite text is used as positive samples, while satellite images and text from other locations are used as negative samples. Cloud-masked low-light enhancement samples are constructed through brightness reduction, local occlusion, fogging perturbation, and contrast reduction, enabling the model to learn stable matching cues under low-light conditions. During the inference phase, only real-time UAV images need to be input to complete low-light perception, enhancement, feature repair, fusion retrieval, and Top-K geographic location output.

[0111] 10. Corresponding solutions to the difficulties in the field

[0112] To address occlusion issues, the system reduces the negative impact of areas obscured by clouds, tree canopies, or shadows through occlusion scoring and textual semantic compensation. To address image distortion and cross-view geometric differences, the system uses triplet spatial relationships and bidirectional visual alignment to maintain structural consistency among roads, buildings, and open areas. To address radiation differences and low-light degradation, the system uses adaptive enhancement and illumination-invariant textual semantic inpainting to avoid noise amplification caused by simple pixel enhancement. For seasonal changes, the system prioritizes matching seasonally stable attributes such as road frameworks, building layouts, water body boundaries, and regional topology.

[0113] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A text-based semantic-guided cross-view geolocation system for low-light UAV images obscured by clouds, characterized in that, include: The drone low-light image input module is used to acquire real-time images collected by the drone under cloud cover conditions. The cloud cover and dim light level perception module is used to calculate the normalized brightness distribution, local contrast, dark area ratio, edge response and fogging index of the real-time image, and calculate the cloud cover dim light score α based on the above indexes. An adaptive low-light enhancement module is used to generate a spatially varying low-light enhancement mask based on the cloud cover low-light score α, and to perform edge-preserving adaptive enhancement of the UAV image. The triplet text description generation module is used to generate triplet text descriptions for satellite images offline. The image encoding module uses a pre-trained CLIP ViT-B image encoder to extract visual features from the enhanced UAV images and satellite images; The text encoding module uses a pre-trained BERT text encoder to extract semantic features from satellite text descriptions; The semantically guided feature repair module is used to generate semantically repaired UAV features based on the cloud cover dark light score α, the UAV visual features before enhancement, the UAV visual features after enhancement, satellite text semantic features, and satellite visual features, through a gating mechanism. The bidirectional image-to-text / image-to-image alignment module is used to perform bidirectional alignment of UAV images to satellite images, satellite images to UAV images, UAV images to satellite text, and satellite text to UAV images. The feature fusion retrieval module is used to fuse the repaired UAV features with satellite text features to generate a UAV query fusion vector, fuse satellite visual features with satellite text features to generate a satellite candidate fusion vector, and calculate the similarity. The contrastive learning training module is used to train the system end-to-end using bidirectional contrastive learning loss.

2. A text-based semantic-guided cross-view geolocation method for low-light UAV images obscured by clouds, implemented using the system described in claim 1, characterized in that... Includes the following steps: Step 1: Acquire real-time images taken by the drone under cloud cover conditions. and a pre-built georeferenced satellite candidate image library, wherein each satellite image in the satellite candidate image library All are associated with pre-generated offline triple text descriptions. ; For the real-time image Calculate its normalized brightness distribution Local contrast normalization value Dark area ratio Normalized edge response and atomization index The cloud cover dark light score is calculated using the following formula. : in, A collection of image pixels. Let σ(·) be the threshold for the dark area, and let σ(·) be the Sigmoid function. , , , , and These are learnable parameters or validated weights; Step 2: Based on the cloud cover dark light score α and the local brightness map, generate the dark light enhancement mask Mu(x) according to the following formula and perform adaptive enhancement on the UAV image: in, G(·) is the edge protection coefficient, G(·) is the edge preservation enhancement function, and clip(·, 0, 1) is the cutoff function. Step 3: Extract the enhanced UAV images using the pre-trained CLIP ViT-B image encoder φI(·). visual features and satellite images visual features The pre-trained BERT text encoder φT(·) is used to extract triplet text descriptions of satellite images. semantic features : This formula maps the UAV visual information, satellite visual information, and illumination-invariant text semantics before and after low-light enhancement to a unified feature space, providing input for subsequent semantic repair and cross-view retrieval; The image encoder and text encoder remain frozen during training; Step 4: Based on the cloud cover low light score α and the enhanced visual features of the UAV... Enhanced visual features of drones Satellite text semantic features and satellite visual features Generate semantically repaired drone features using the following formula. : Where [·;·] represents feature concatenation, and Here are the parameters of the gated network, and ψ(·) represents the semantic correction network. This represents element-wise multiplication, and Norm(·) is the normalization function. Step 5: Extract the semantically repaired drone features With satellite text semantic features The concatenated vectors are input into a multilayer perceptron to generate a UAV query fusion vector. Satellite visual features With satellite text semantic features The concatenated data is then input into a multilayer perceptron to generate satellite candidate fusion vectors. ; The similarity between the UAV query fusion vector and the candidate satellite fusion vectors is calculated using the following formula. in, For drone query fusion vectors, For satellite candidate fusion vectors; Step Six: Based on similarity The satellite candidate image library is sorted from highest to lowest, and the top-K satellite candidate images and their corresponding geographic coordinates are output.

3. The text semantic-guided cross-view geolocation method for low-light UAV images obscured by clouds, as described in claim 2, is characterized in that: The triplet text description The generation method is as follows: For each satellite candidate image, the multimodal large language model Qwen2-VL is called offline to first generate attribute prompts, and then the attribute prompts are input into the multimodal large language model to generate fine-grained text descriptions; the attribute prompts include building layout, road layout and building style; the fine-grained text descriptions include building shape, road direction, spatial relationship and building appearance; the generated prompts explicitly require priority to describe attributes that are constant in illumination and relatively stable in season, including road framework, building layout, roof outline, water body boundary, playground shape, open area and vegetation edge.

4. The text semantic-guided cross-view geolocation method for low-light UAV images obscured by clouds, as described in claim 3, is characterized in that: The edge-preserving enhancement function G(·) is implemented using guided filtering or bilateral filtering.

5. A text-based semantic-guided cross-view geolocation method for low-light UAV images obscured by clouds, as described in claim 4, is characterized in that: The semantic correction network ψ(·) is implemented using a cross-attention mechanism to enhance the features of the post-drone. As a query, using satellite text semantic features and satellite visual features The concatenation of values ​​serves as the key and value, and semantically enhanced drone features are obtained through multi-head cross-attention computation.

6. A text-based semantic-guided cross-view geolocation method for low-light UAV images obscured by clouds, as described in claim 5, is characterized in that: The following loss function is used for optimization during the training phase: in, For the two-way alignment loss between UAV imagery and satellite imagery, For the semantic alignment loss of UAV image-satellite text, To repair the consistency constraint loss for features, Classify losses for optional locations. For temperature coefficient, , , , Let P(·) be the loss weight and P(·) be the projection function.

7. A text-based semantic-guided cross-view geolocation method for low-light UAV images obscured by clouds, as described in claim 6, is characterized in that: During the training phase, cloud-masked dark light enhancement samples are constructed by reducing brightness, local occlusion, fogging perturbation, and contrast reduction.

Images