Foggy day cross-view-angle geographic positioning method and device, computer equipment and storage medium

Abstract

The invention relates to a foggy day cross-view geographic positioning method and device, computer equipment and a storage medium. The method comprises the following steps: inputting a to-be-positioned image and an image library image set as input images into a pre-trained cross-view-angle positioning model, and determining compact semantic features corresponding to the input images; the cross-view positioning model comprises an adaptive feature defogging layer and a dual-path feature enhancement layer; specifically, the adaptive feature defogging layer performs multi-scale decomposition, adaptive filtering, detail guide fusion contrast enhancement and detail enhancement processing on initial features of an input image to obtain defogging layer output features; based on a double-path feature enhancement layer, performing double-path feature enhancement on the output features of the defogging layer to obtain compact semantic features; and based on the to-be-positioned image and the compact semantic features corresponding to the image library images in the image library image set, determining a target matching image corresponding to the to-be-positioned image in the image library image set. By adopting the method, the cross-view-angle positioning precision can be improved.

Description

Technical Field

[0001] This application relates to the field of computer vision technology, and in particular to a method, apparatus, computer device, and storage medium for cross-view geolocation in foggy weather. Background Technology

[0002] With the acceleration of global urbanization and the widespread application of intelligent transportation systems, high-precision geolocation technology has become a core support for many fields such as autonomous driving, drone navigation, and drone delivery. Cross-view geolocation aims to determine the geographical location information of a target image by matching images of the same location from different perspectives (such as drones and satellites). This technology does not rely on the assistance of the Global Navigation Satellite System (GNSS) and shows unique advantages in scenarios where GNSS signals are limited or unavailable, such as urban canyons.

[0003] In real-world applications, complex weather conditions such as fog can impact image quality. For detailed UAV images, fog blurs crucial local textures and structural information, diminishing their advantage in high-resolution detail. For satellite images with wide coverage but limited resolution, fog reduces overall contrast and sharpness, making already limited discriminative features even more difficult to extract and match, increasing the difficulty of cross-view matching.

[0004] Existing cross-view geolocation methods are mainly based on CNN (Convolutional Neural Network) or Transformer. These methods can achieve high positioning accuracy under ideal weather conditions, but under adverse weather conditions such as fog, images will suffer from problems such as decreased contrast, color distortion, and blurred texture details, which makes feature extraction and matching difficult and significantly reduces positioning accuracy. Summary of the Invention

[0005] Therefore, it is necessary to provide a foggy weather cross-view geolocation method, device, computer equipment, and storage medium that can solve the problem of feature extraction and matching difficulties and thus improve cross-view positioning accuracy.

[0006] Firstly, this application provides a cross-view geolocation method for foggy weather, including:

[0007] S1 takes the image to be located and the image set in the image library as input images, inputs them into the pre-trained cross-view localization model, and determines the compact semantic features corresponding to each image in the image library and the image set in the image library respectively; the cross-view localization model includes an adaptive feature dehazing layer and a dual-path feature enhancement layer;

[0008] The cross-view localization model processes the input image as follows:

[0009] An adaptive feature dehazing layer is used to perform multi-scale decomposition and adaptive filtering on the initial features corresponding to the input image to obtain enhanced base features. Then, guided detail fusion is performed on the enhanced base features to obtain refined features. and the characteristics after refinement Contrast enhancement and detail enhancement processes are performed to obtain the output features of the dehazing layer. ;

[0010] Based on the dual-path feature enhancement layer, the output features of the dehazing layer are... Channel path enhancement and spatial adaptive enhancement are performed to obtain enhanced features, and then compressed projection processing is applied to the enhanced features to obtain compact semantic features corresponding to the input image. ;

[0011] S2, based on the image to be located and the compact semantic features corresponding to each image in the image library, determine the target matching image corresponding to the image to be located in the image library.

[0012] In one embodiment, the process of performing multi-scale decomposition processing on the initial features corresponding to the input image based on the adaptive feature dehazing layer includes:

[0013] Initial features corresponding to the input image Using a step size of 2 Convolution is used to process the data and obtain the basic low-frequency features. and initial features Using a step size of 1 Convolution is used to process the data to obtain features of the same scale. Initial features The initial features were obtained by using DINOv2 to extract features from the input image, where the input image is a dataset of foggy images.

[0014] Basic low-frequency characteristics Perform bilinear upsampling and then Convolution adjustment yields reconstructed base layer features Reconstructing the characteristics of the base layer The calculation formula is ,in, Bilinear upsampling;

[0015] Calculate features at the same scale With reconstruction of basic layer features The residuals between them yield detailed features. , For detailed features.

[0016] In one embodiment, the adaptive filtering process includes:

[0017] Based on initial features Initial features are generated using two consecutive convolutions. The corresponding space-channel adaptive weight matrix is , For the Sigmoid function;

[0018] Space-channel adaptive weight matrix With reconstruction of basic layer features Element-wise multiplication yields the enhanced basic features. , The symbol for element-wise multiplication.

[0019] In one embodiment, guided detail fusion is performed on the enhanced basic features to obtain refined features. ,include:

[0020] Enhanced basic features With detailed features Perform channel-dimensional concatenation to obtain the fused features. ;

[0021] Fusion features After performing convolution and normalization, the refined features are obtained. .

[0022] In one embodiment, the refined features Contrast enhancement and detail enhancement processes are performed to obtain the output features of the dehazing layer. ,include:

[0023] Characteristics after refining Global average pooling and convolution are performed to obtain refined features. The channel importance weight is And assign channel importance weights Characteristics acting on refined products The contrast enhancement features obtained are ;in, Based on the characteristics of the refined The result is obtained by global average pooling, and the calculation formula is as follows: ;

[0024] Enhance contrast and strengthen features through detailed enhancement processing. The high-frequency detail information is used to obtain detail enhancement features. ;

[0025] Based on detail enhancement features and initial features Determine the output characteristics of the defogging layer as follows .

[0026] In one embodiment, the dehazing layer outputs features based on a dual-path feature enhancement layer. Channel path enhancement and spatial adaptive enhancement are performed to obtain enhanced features, and then compressed projection processing is applied to the enhanced features to obtain compact semantic features corresponding to the input image. ,include:

[0027] Output characteristics of the dehazing layer Convolution and nonlinear transformation processes are performed to obtain the output features of the dehazing layer. The importance weights of each channel are: and output features for the dehazing layer. Convolution and normalization are performed to obtain the output features of the dehazing layer. The spatial enhancement coefficient is In the formula, For the Sigmoid function;

[0028] Employing a learnable gating network, the output features of the dehazing layer are fused. Importance weights of each channel and spatial enhancement coefficient The fusion enhancement coefficient is obtained as ,in, , From the spatial enhancement coefficient This is obtained by expanding along the channel dimension to the original number of channels;

[0029] Based on fusion enhancement coefficient and the output characteristics of the defogging layer The enhanced features are For the enhanced features Compressed projection processing is performed to obtain compact semantic features corresponding to the input image. .

[0030] In one embodiment, the pre-training process of the cross-view localization model includes:

[0031] Obtain the training dataset; the training dataset includes query images and a gallery image set, which are cross-view images.

[0032] A cross-view localization model is constructed based on a pre-defined cross-view localization network and a training dataset; the cross-view localization network includes an adaptive feature dehazing layer and a dual-path feature enhancement layer;

[0033] Compact semantic features corresponding to the query image are determined based on an adaptive feature dehazing layer and a dual-path feature enhancement layer. Compact semantic features corresponding to each image in the image library set. And the compact semantic features of the positive samples corresponding to the query image. ;

[0034] Based on the compact semantic features corresponding to the query image Compact semantic features corresponding to each image in the image library set. And the compact semantic features of the positive samples corresponding to the query image. The loss function is determined as follows: In the formula, For the first image in the image library collection Compact semantic features corresponding to the images For temperature parameters;

[0035] The cross-view localization network is optimized using a loss function to obtain a cross-view localization model.

[0036] Secondly, this application also provides a foggy weather cross-view geolocation device, comprising:

[0037] The feature extraction module is used to take the image to be located and the image set in the image library as input images, input them into the pre-trained cross-view localization model, and determine the compact semantic features corresponding to the image to be located and each image in the image library; the cross-view localization model includes an adaptive feature dehazing layer and a dual-path feature enhancement layer;

[0038] The cross-view localization model processes the input image as follows:

[0039] An adaptive feature dehazing layer is used to perform multi-scale decomposition and adaptive filtering on the initial features corresponding to the input image to obtain enhanced base features. Then, guided detail fusion is performed on the enhanced base features to obtain refined features. and the characteristics after refinement Contrast enhancement and detail enhancement processes are performed to obtain the output features of the dehazing layer. ;

[0040] Based on the dual-path feature enhancement layer, the output features of the dehazing layer are... Channel path enhancement and spatial adaptive enhancement are performed to obtain enhanced features, and then compressed projection processing is applied to the enhanced features to obtain compact semantic features corresponding to the input image. ;

[0041] The localization module is used to determine the target matching image corresponding to the image to be localized in the image library based on the image to be localized and the compact semantic features corresponding to each image in the image library.

[0042] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0043] S1 takes the image to be located and the image set in the image library as input images, inputs them into the pre-trained cross-view localization model, and determines the compact semantic features corresponding to each image in the image library and the image set in the image library respectively; the cross-view localization model includes an adaptive feature dehazing layer and a dual-path feature enhancement layer;

[0044] The cross-view localization model processes the input image as follows:

[0045] An adaptive feature dehazing layer is used to perform multi-scale decomposition and adaptive filtering on the initial features corresponding to the input image to obtain enhanced base features. Then, guided detail fusion is performed on the enhanced base features to obtain refined features. and the characteristics after refinement Contrast enhancement and detail enhancement processes are performed to obtain the output features of the dehazing layer. ;

[0046] Based on the dual-path feature enhancement layer, the output features of the dehazing layer are... Channel path enhancement and spatial adaptive enhancement are performed to obtain enhanced features, and then compressed projection processing is applied to the enhanced features to obtain compact semantic features corresponding to the input image. ;

[0047] S2, based on the image to be located and the compact semantic features corresponding to each image in the image library, determine the target matching image corresponding to the image to be located in the image library.

[0048] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the following steps:

[0049] S1 takes the image to be located and the image set in the image library as input images, inputs them into the pre-trained cross-view localization model, and determines the compact semantic features corresponding to each image in the image library and the image set in the image library respectively; the cross-view localization model includes an adaptive feature dehazing layer and a dual-path feature enhancement layer;

[0050] The cross-view localization model processes the input image as follows:

[0051] An adaptive feature dehazing layer is used to perform multi-scale decomposition and adaptive filtering on the initial features corresponding to the input image to obtain enhanced base features. Then, guided detail fusion is performed on the enhanced base features to obtain refined features. and the characteristics after refinement Contrast enhancement and detail enhancement processes are performed to obtain the output features of the dehazing layer. ;

[0052] Based on the dual-path feature enhancement layer, the output features of the dehazing layer are... Channel path enhancement and spatial adaptive enhancement are performed to obtain enhanced features, and then compressed projection processing is applied to the enhanced features to obtain compact semantic features corresponding to the input image. ;

[0053] S2, based on the image to be located and the compact semantic features corresponding to each image in the image library, determine the target matching image corresponding to the image to be located in the image library.

[0054] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, performs the following steps:

[0055] S1 takes the image to be located and the image set in the image library as input images, inputs them into the pre-trained cross-view localization model, and determines the compact semantic features corresponding to each image in the image library and the image set in the image library respectively; the cross-view localization model includes an adaptive feature dehazing layer and a dual-path feature enhancement layer;

[0056] The cross-view localization model processes the input image as follows:

[0057] An adaptive feature dehazing layer is used to perform multi-scale decomposition and adaptive filtering on the initial features corresponding to the input image to obtain enhanced base features. Then, guided detail fusion is performed on the enhanced base features to obtain refined features. and the characteristics after refinement Contrast enhancement and detail enhancement processes are performed to obtain the output features of the dehazing layer. ;

[0058] Based on the dual-path feature enhancement layer, the output features of the dehazing layer are... Channel path enhancement and spatial adaptive enhancement are performed to obtain enhanced features, and then compressed projection processing is applied to the enhanced features to obtain compact semantic features corresponding to the input image. ;

[0059] S2, based on the image to be located and the compact semantic features corresponding to each image in the image library, determine the target matching image corresponding to the image to be located in the image library.

[0060] The aforementioned foggy cross-view geolocation method, device, computer equipment, and storage medium include the following steps: S1, taking the image to be located and a set of images from a library as input images, inputting them into a pre-trained cross-view geolocation model, and determining the compact semantic features corresponding to each image in the image library and the image to be located; the cross-view geolocation model includes an adaptive feature dehazing layer and a dual-path feature enhancement layer; the cross-view geolocation model's processing of the input image includes: performing multi-scale decomposition and adaptive filtering on the initial features corresponding to the input image based on the adaptive feature dehazing layer to obtain enhanced basic features, and performing detail-guided fusion on the enhanced basic features to obtain refined features. and the characteristics after refinement Contrast enhancement and detail enhancement processes are performed to obtain the output features of the dehazing layer. Based on the dual-path feature enhancement layer, the output features of the dehazing layer are improved. Channel path enhancement and spatial adaptive enhancement are performed to obtain enhanced features, and then compressed projection processing is applied to the enhanced features to obtain compact semantic features corresponding to the input image. S2, based on the compact semantic features corresponding to the image to be located and each image in the image library, determine the target matching image corresponding to the image to be located in the image library. This invention addresses the problem of decreased accuracy in cross-view geolocation under foggy conditions by proposing a cross-view geolocation method for foggy weather. An adaptive feature defogging layer is designed, which directly performs adaptive enhancement and purification at the feature level, thereby suppressing noise and degradation caused by fog, while retaining and enhancing structural and detailed information crucial to the geolocation task. To further improve the discriminative power of features, a dual-path feature enhancement layer is designed. This dual-path feature enhancement layer evaluates the importance of features in parallel from both channel and spatial dimensions, and by fusing the outputs of the two paths, achieves refined focusing on key areas and suppression of redundant information, thereby generating a more robust feature representation. Ultimately, this solves the problem of difficult feature extraction and matching, and improves the accuracy of cross-view geolocation in foggy environments. Attached Figure Description

[0061] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies 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.

[0062] Figure 1 This is an application environment diagram of a cross-view geolocation method in foggy weather in one embodiment;

[0063] Figure 2This is a flowchart illustrating the process of constructing a cross-view localization model in one embodiment;

[0064] Figure 3 This is a schematic diagram of the adaptive feature dehazing layer in one embodiment;

[0065] Figure 4 This is a flowchart illustrating the dual-path feature enhancement layer in one embodiment;

[0066] Figure 5 This is a structural block diagram of a cross-view geolocation device in foggy weather, as shown in one embodiment. Detailed Implementation

[0067] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0068] In one exemplary embodiment, such as Figure 1 As shown, a cross-view geolocation method for foggy weather is provided, which includes the following steps 102 to 104.

[0069] Step 102: The image to be located and the image set in the image library are used as input images. The pre-trained cross-view localization model is input to determine the compact semantic features corresponding to each image in the image library and the image to be located. The cross-view localization model includes an adaptive feature dehazing layer and a dual-path feature enhancement layer. The processing of the input image by the cross-view localization model includes: performing multi-scale decomposition and adaptive filtering on the initial features corresponding to the input image based on the adaptive feature dehazing layer to obtain enhanced basic features; performing detail-guided fusion on the enhanced basic features to obtain refined features; and performing contrast enhancement and detail enhancement on the refined features to obtain the output features of the dehazing layer. Based on the dual-path feature enhancement layer, performing channel path enhancement and spatial adaptive enhancement on the output features of the dehazing layer to obtain enhanced features; and performing compression projection processing on the enhanced features to obtain the compact semantic features corresponding to the input image.

[0070] For example, such as Figure 2 The diagram shows the process flow of the cross-view localization model for processing the input image. The process includes the following steps:

[0071] Step 1022: Based on the adaptive feature dehazing layer, the initial features corresponding to the input image are decomposed into multiple scales and subjected to adaptive filtering to obtain enhanced basic features. Then, guided detail fusion is performed on the enhanced basic features to obtain refined features. and the characteristics after refinement Contrast enhancement and detail enhancement processes are performed to obtain the output features of the dehazing layer. .

[0072] For example, the initial features corresponding to the input image are fed into the adaptive feature dehazing layer, such as... Figure 3 The diagram illustrates the workflow of the adaptive feature dehazing layer. This layer first performs multi-scale decomposition on the initial input features, extracting a basic low-frequency layer through downsampling convolution, preserving spatial structure through same-scale convolution, and then obtaining a high-frequency detail layer through upsampling reconstruction and residual calculation. Next, adaptive filtering is applied to the basic layer, learning a spatial-channel weight matrix to enhance edges and suppress noise. Then, a detail-guided fusion mechanism is used to concatenate the enhanced basic layer and the detail layer, reconstructing the fog-covered detail structure through convolutional recombination and normalization. A global contrast enhancement mechanism is then employed, learning enhancement weights based on channel statistics to improve feature discriminative power in the form of residuals. Finally, high-frequency detail enhancement is further strengthened through convolutional layers, and residual connections are made with the original input to obtain the dehazing layer output features. This enhances features while preserving the integrity of the original information.

[0073] Step 1024: Based on the dual-path feature enhancement layer, output features from the dehazing layer. Channel path enhancement and spatial adaptive enhancement are performed to obtain enhanced features, and then compressed projection processing is applied to the enhanced features to obtain compact semantic features corresponding to the input image. .

[0074] The dehazing layer output features obtained in step 1042 Feed it into the dual-path feature enhancement layer, such as Figure 4 The diagram shown illustrates the workflow of the dual-path feature enhancement layer. First, the input dehazing layer outputs features... The inputs are fed into two parallel paths: a channel enhancement path and a spatial adaptive enhancement path. The channel enhancement path learns the importance weights of each channel through a bottleneck structure, highlighting discriminative feature responses. The spatial adaptive enhancement path captures spatial context dependencies and generates spatially sensitive enhancement coefficients. The outputs of these two parallel paths are then concatenated, and adaptive gating weights are learned through convolution and the Softmax function to dynamically balance the contribution ratio of channel and spatial information. Finally, the learned gating weights are used to weight and fuse the outputs of the two parallel paths, and the fused weights are applied to the original features to obtain enhanced features, achieving refined focusing on key regions and effective suppression of redundant information.

[0075] Step 104: Based on the image to be located and the compact semantic features corresponding to each image in the image library, determine the target matching image corresponding to the image to be located in the image library.

[0076] Optionally, by calculating the cosine similarity between the compact semantic features of the image to be located and all images in the image library set, a series of similarity scores are obtained. Based on these similarity scores, all images in the image library set are sorted in descending order, and the image with the highest score is taken as the best match for the image to be located, i.e., the target matching image corresponding to the image to be located. This enables accurate cross-view geolocation. In this invention, the entire process, from initial feature extraction to similarity calculation, forms a complete closed loop, significantly improving positioning performance under foggy conditions.

[0077] This invention addresses the issue of decreased accuracy in cross-view geolocation under foggy conditions by proposing a method for cross-view geolocation in foggy weather. An adaptive feature defogging layer is designed, which directly enhances and purifies features at the feature level, thereby suppressing noise and degradation caused by fog while preserving and enhancing structural and detailed information crucial for geolocation. To further improve feature discriminative power, a dual-path feature enhancement layer is designed. This layer evaluates feature importance in parallel from both channel and spatial dimensions, and by fusing the outputs of the two paths, it achieves refined focusing on key areas and suppression of redundant information, thus generating more robust feature representations. Ultimately, this method solves the problems of feature extraction and matching difficulties and improves the accuracy of cross-view geolocation in foggy environments.

[0078] In one exemplary embodiment, please refer again Figure 3 Step 1022 includes the following steps.

[0079] Step (a), for the initial features Using a step size of 2 Convolution is used to process the data and obtain the basic low-frequency features. and initial features Using a step size of 1 Convolution is used to process the data to obtain features of the same scale. Initial features The initial features were obtained by using DINOv2 to extract features from the input image, which is a dataset of foggy images; basic low-frequency features were extracted. Perform bilinear upsampling and then Convolution adjustment yields reconstructed base layer features Reconstructing the characteristics of the base layer The calculation formula is ,in, Bilinear upsampling; calculation of features at the same scale With reconstruction of basic layer features The residuals between them yield detailed features. , For detailed features.

[0080] Optionally, this invention uses DINOv2, trained on a large-scale dataset, as the basic feature extractor to extract features from the input image (i.e., satellite images and UAV images from multiple perspectives), generating a high-dimensional initial feature representation, and thus obtaining the initial features corresponding to the input image. .

[0081] To address the issues of reduced image contrast and loss of detail under foggy conditions, this invention proposes a lightweight adaptive feature dehazing layer. This layer does not rely on traditional atmospheric scattering models but instead combines traditional image priors with deep learning to restore and enhance degraded features in the feature space. Fog interference exhibits differentiated characteristics across different scales; for example, low-frequency regions are significantly affected by fog, resulting in an overall decrease in contrast, while high-frequency regions such as edges and textures exhibit localized detail blurring. To specifically address these effects, this invention utilizes an adaptive feature dehazing layer to first process the initial input features... Perform multi-scale decomposition.

[0082] Specifically, regarding the initial features Perform a step of 2. Convolutional downsampling processing is used to obtain basic low-frequency features such as large-scale structural information and color distribution. At the same time, for the initial features Perform a step size of 1 Convolution obtains features of the same scale Preserving the original spatial geometry. Basic low-frequency characteristics. and features of the same scale The formula for determining it is:

[0083]

[0084] In the above formula, It has a step size of 2. convolution, It has a step size of 1. Convolution. For basic low-frequency features. Perform bilinear upsampling and pass through Convolution adjustment yields reconstructed base layer features Finally, by calculating features at the same scale... With reconstruction of basic layer features The residuals between them are used to obtain detailed features. Detailed features It includes high-frequency information such as edges and textures from the initial features, providing a foundation for subsequent detail enhancement. This includes reconstructing the base layer features. and detailed features The formula is determined as follows:

[0085]

[0086] In the above formula, It is bilinear upsampling.

[0087] Step (b), based on initial features Initial features are generated using two consecutive convolutions. The corresponding space-channel adaptive weight matrix is ; adapt the spatial-channel weight matrix With reconstruction of basic layer features Element-wise multiplication yields the enhanced basic features. .

[0088] The initial feature base layer contains a large amount of low-frequency information, which is severely affected by fog. To enhance the feature representation of these regions, this paper performs adaptive filtering. Specifically, through two consecutive... Convolution learns initial features The corresponding spatial-channel adaptive weight matrix is ​​then used to reconstruct the base layer features. Element-wise multiplication yields the enhanced basic features. This allows the network to adaptively enhance the response of edge and textured regions while smoothing flat regions, achieving edge-preserving and noise reduction. The expression for the above process is shown below:

[0089]

[0090] In the above formula, For the Sigmoid function, This is for element-wise multiplication.

[0091] Step (c) involves enhancing the basic features. With detailed features Perform channel-dimensional concatenation to obtain the fused features. ; fusion features After performing convolution and normalization, the refined features are obtained. .

[0092] The features of the detail layer of the initial features contain relatively complete high-frequency information, which can serve as guiding signals to promote the detail recovery of its base layer. Specifically, using a detail-guided fusion mechanism, the enhanced base features are... With detailed features Channel-dimensional splicing is performed to obtain fused features. Fusion features First pass Convolution is used for feature recombination, then normalization is used to eliminate scale shift caused by differences in fog concentration, followed by ReLU activation, and then... Convolution yields refined features This process enables the effective propagation of detailed information to the base layer, reconstructing the detailed structure obscured by fog. The expression for the above process is shown below:

[0093]

[0094] Step (d) refines the characteristics Global average pooling and convolution are performed to obtain refined features. The channel importance weight is And assign channel importance weights Characteristics acting on refined products The contrast enhancement features obtained are ;in, Based on the characteristics of the refined The result is obtained by global average pooling, and the calculation formula is as follows: .

[0095] To compensate for the decrease in feature contrast caused by fog, this invention introduces a contrast enhancement mechanism based on global feature statistics. Firstly, the refined features... Perform global average pooling to obtain channel-level global statistics, then process them through two... Convolutional layer learns channel importance weights This weight is applied to the refined features in the form of residual scaling. By specifically enhancing channels with high discriminative power, contrast enhancement features are obtained. The expression is as follows:

[0096]

[0097] in, Based on the characteristics of the refined The result is obtained by performing global average pooling. For the Sigmoid function, This is for element-wise multiplication.

[0098] Step (e) enhances contrast and features through detail enhancement processing. The high-frequency detail information is used to obtain detail enhancement features. ;Detail-enhanced features and initial features Determine the output characteristics of the defogging layer as follows .

[0099] After the aforementioned processing, the overall structure of the initial features has been largely restored, but local details still need further enhancement. Therefore, a detail enhancement mechanism is introduced, which uses... Convolution further enhances high-frequency detail information, employing the idea of ​​unsharpening masking to enhance texture features, as shown in the following expression:

[0100]

[0101] Finally, residual connections are used to preserve the original information of the initial features, and the output features of the dehazing layer are determined as follows: .

[0102] In this embodiment, the entire processing flow of the adaptive feature dehazing layer is completed entirely in the feature space, without relying on physical model parameter estimation. This effectively improves the feature quality under foggy conditions, achieves robust end-to-end feature learning, and provides more reliable feature representations for subsequent cross-view matching tasks. Specifically, addressing the image degradation problem caused by fog, this invention innovatively designs an adaptive feature dehazing layer. This layer performs adaptive enhancement in the feature space by fusing multi-scale decomposition, adaptive filtering, and detail-guided fusion mechanisms, effectively restoring the contrast and detail information of foggy images.

[0103] In one exemplary embodiment, please refer again Figure 4 Step 1024 includes: outputting features for the dehazing layer. Convolution and nonlinear transformation processes are performed to obtain the output features of the dehazing layer. The importance weights of each channel are: and output features for the dehazing layer. Convolution and normalization are performed to obtain the output features of the dehazing layer. The spatial enhancement coefficient is In the formula, The function is a sigmoid function; a learnable gating network is used to fuse the output features of the dehazing layer. Importance weights of each channel and spatial enhancement coefficient The fusion enhancement coefficient is obtained as ,in, , From the spatial enhancement coefficient Obtained by extending along the channel dimension to the original number of channels; based on the fusion enhancement coefficient. and the output characteristics of the defogging layer The enhanced features are For the enhanced features Compressed projection processing is performed to obtain compact semantic features corresponding to the input image. .

[0104] While traditional dehazing modules can effectively improve feature contrast and overall clarity, they have two inherent limitations: first, filtering operations may smooth out details, weakening subtle but crucial texture features; second, globally uniform enhancement is difficult to adapt to the differentiated enhancement needs of different regions. To address these issues, this invention designs a dual-path feature enhancement layer. This dual-path feature enhancement layer serves as a post-processing unit of the adaptive feature dehazing layer. By finely recalibrating the dehazed features output from the previous stage, it effectively compensates for detail loss and achieves region-adaptive feature enhancement, significantly improving visual quality and semantic integrity.

[0105] Specifically, to compensate for the detail smoothing problem caused by adaptive filtering, the channel enhancement path in the dual-path feature enhancement layer adopts a bottleneck structure design, independently learning the importance weights of each channel at each position. First, through... After convolution compresses the channel dimension, a non-linear transformation is introduced using the GELU activation function, followed by... Convolution is used to restore the original channel dimensions, and finally, the sigmoid function is used to generate the importance weights of each channel at each position. Importance weight of each channel It can effectively sharpen subtle textures and high-frequency features that have been weakened by filtering operations, as expressed by: .

[0106] To address the issue of global augmentation failing to differentiate between regions, the spatially adaptive augmentation path captures local contextual relationships to generate spatially sensitive feature augmentation coefficients. Specifically, it first... Convolution captures spatial dependencies, followed by normalization and ReLU activation, and then... Convolution compresses the number of channels to 1, and finally, spatial enhancement coefficients are generated using the Sigmoid function. The expression is Spatial enhancement factor Expand along the channel dimension to the original number of channels Later obtained It is used to enhance structural response in areas with complex edges and textures, and to suppress noise in flat areas.

[0107] Furthermore, this paper introduces a learnable gating network to dynamically fuse the importance weights of each channel. and spatial enhancement coefficient The fusion enhancement coefficient was obtained. This enables precise region adaptive processing, resulting in enhanced features. The enhanced features While preserving the basic information of dehazing, the details are enhanced and the different enhancement needs of different regions are adapted, ultimately outputting enhanced features with higher visual quality and more complete semantic information. The expression of the above process is as follows:

[0108]

[0109] In the above formula, This is a gate value that automatically determines the needs of each region. In areas with rich texture, it favors channel enhancement paths to strengthen feature responses, while in structural edge areas, it favors spatially adaptive paths to maintain contour clarity and geometric continuity. Based on this gate value... Filtration enhancement coefficient obtained Based on fusion enhancement coefficient and the output characteristics of the defogging layer Enhanced features Enhanced features through Convolution compresses projection to obtain compact semantic features. .

[0110] In this embodiment, the dehazing layer output feature obtained after adaptive feature dehazing layer enhancement is... The data is fed into a dual-path feature enhancement layer. This layer uses parallel channel enhancement paths and spatial adaptive enhancement paths to enhance the output features of the dehazing layer from both the feature channel and spatial structure dimensions, respectively. By performing refined recalibration and dynamically fusing the outputs of the two paths using a learnable gating network, the discriminativeness and robustness of the features are further improved, thereby enhancing cross-view positioning accuracy.

[0111] In an exemplary embodiment, the pre-training process of the cross-view localization model includes: acquiring a training dataset; the training dataset includes a query image and a library image set, wherein the query image and the library image set are cross-view images; constructing a cross-view localization model based on a pre-defined cross-view localization network and the training dataset; the cross-view localization network includes an adaptive feature dehazing layer and a dual-path feature enhancement layer; and determining compact semantic features corresponding to the query image based on the adaptive feature dehazing layer and the dual-path feature enhancement layer. Compact semantic features corresponding to each image in the image library set. And the compact semantic features of the positive samples corresponding to the query image. Based on the compact semantic features corresponding to the query image Compact semantic features corresponding to each image in the image library set. And the compact semantic features of the positive samples corresponding to the query image. The loss function is determined as follows: In the formula, For the first image in the image library collection Compact semantic features corresponding to the images The temperature parameter is used; the cross-view localization network is optimized using the loss function to obtain the cross-view localization model.

[0112] Obtain a training dataset for training the cross-view localization model. This training dataset includes query images and corresponding image sets from a gallery. The query images and the images in the gallery are images of the same geographic area from different perspectives (such as drones, satellites, etc.). The query images are used to simulate the target images to be matched in actual cross-view localization, while the image sets are used to provide candidate matching images.

[0113] Optionally, this invention uses the University-1652 dataset to perform dual-view consistent fogging processing on UAV and satellite images, constructing the Fog-University dataset to simulate foggy environments, thus obtaining the training dataset. The University-1652 dataset is a multi-view, cross-platform dataset designed for UAV geolocation tasks, integrating satellite, UAV, and ground-view images. Compared to other datasets that only select prominent landmarks as targets, University-1652 innovatively selects 1652 ordinary buildings from 72 universities worldwide as identification objects. On average, each building in this dataset contains one satellite image and 54 UAV images. In terms of data partitioning, the University-1652 dataset adopts a non-overlapping partitioning strategy. The training set consists of 701 buildings from 33 universities, while the test set consists of 951 buildings from another 39 universities, ensuring the objectivity and validity of the evaluation results. The Fog-University dataset is derived from the University-1652 dataset by applying dual-view consistent fogging to UAV and satellite images. Its purpose is to simulate the impact of fog conditions on cross-view geolocation observations while avoiding modal bias caused by single-view fogging. Similar to the University-1652 dataset, the Fog-University dataset supports two main tasks: 1) Fog-fighting UAV localization: Given an image taken from a foggy UAV's perspective, find matching foggy satellite images. Specifically, using 37,854 fogged UAV images as the query set, matching is performed in a library containing 701 fogged satellite images and 250 interference items; 2) Fog-fighting UAV navigation: Given a foggy satellite image, retrieve matching foggy UAV images. Specifically, using 701 fogged satellite images as the query set, target matching is achieved in a library containing 37,854 fogged UAV images and 13,501 interference items.

[0114] During the training of the cross-view localization model, the symmetric InfoNCE loss function is used to optimize the feature space, as shown in the following expression:

[0115]

[0116] In the above formula, The compact semantic features corresponding to the query image determined through the above process. For the first image in the image library collection The compact semantic features corresponding to the images, where R is the number of images in the image library set. It is the only positive sample that matches the query image q. For temperature parameters. When querying image q and positive samples When similar, the loss value decreases; conversely, when dissimilar, the loss value increases.

[0117] In this embodiment, the symmetric InfoNCE loss function is used to optimize the cross-view localization network to obtain a cross-view localization model, which enables the cross-view localization model to learn cross-view feature representations that are insensitive to fog conditions. The entire framework adopts an end-to-end training method, which significantly improves the performance of the cross-view localization model in fog conditions while maintaining high localization accuracy.

[0118] Experiments on the University-1652 dataset show that the present invention achieves an R@1 of 95.75% and an AP of 96.30% in UAV localization tasks; and an R@1 of 96.14% and an AP of 95.44% in UAV navigation tasks, comparable to the current state-of-the-art CV-cities model. On the self-built foggy dataset Fog-University, the present invention achieves an AP of 96.21% in foggy UAV localization tasks, a 2.77% improvement over the current state-of-the-art CV-cities model; and an AP of 95.44% in foggy UAV navigation tasks, a 2.24% improvement over the current state-of-the-art CV-cities model. The data demonstrates that the present invention achieves high-precision localization. Specifically, the present invention effectively suppresses feature degradation caused by fog through the synergistic effect of a multi-scale adaptive feature defogging module and a dual-path feature enhancement module, maintaining stable localization performance even in foggy conditions, demonstrating robustness. By embedding prior knowledge such as multi-scale decomposition and filtering from traditional image processing into the deep feature learning process in a differentiable manner, the network computation process is endowed with explicit image processing semantics, resulting in strong interpretability. This avoids separate image preprocessing steps and achieves end-to-end collaborative optimization of the dehazing process and geolocation task, improving the system's practicality and efficiency.

[0119] It should be understood that although the steps in the flowcharts of the above embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0120] Based on the same inventive concept, this application also provides a foggy cross-view geolocation device for implementing the aforementioned foggy cross-view geolocation method. The solution provided by this device is similar to the solution described in the above method; therefore, the specific limitations of the one or more foggy cross-view geolocation device embodiments provided below can be found in the limitations of the foggy cross-view geolocation method described above, and will not be repeated here.

[0121] In one exemplary embodiment, such as Figure 5 As shown, a foggy weather cross-view geolocation device is provided, including: a feature extraction module 502 and a positioning module 504, wherein:

[0122] The feature extraction module 502 is used to take the image to be located and the image set of the image library as input images, input them into the pre-trained cross-view localization model, and determine the compact semantic features corresponding to each image in the image library and the image set of the image library respectively; the cross-view localization model includes an adaptive feature dehazing layer and a dual-path feature enhancement layer;

[0123] The cross-view localization model processes the input image as follows:

[0124] An adaptive feature dehazing layer is used to perform multi-scale decomposition and adaptive filtering on the initial features corresponding to the input image to obtain enhanced base features. Then, guided detail fusion is performed on the enhanced base features to obtain refined features. and the characteristics after refinement Contrast enhancement and detail enhancement processes are performed to obtain the output features of the dehazing layer. ;

[0125] Based on the dual-path feature enhancement layer, the output features of the dehazing layer are... Channel path enhancement and spatial adaptive enhancement are performed to obtain enhanced features, and then compressed projection processing is applied to the enhanced features to obtain compact semantic features corresponding to the input image. .

[0126] The localization module 504 is used to determine the target matching image corresponding to the image to be located in the image library based on the image to be located and the compact semantic features corresponding to each image in the image library.

[0127] In an exemplary embodiment, the feature extraction module 502 is further configured to extract initial features corresponding to the input image. Using a step size of 2 Convolution is used to process the data and obtain the basic low-frequency features. and initial features Using a step size of 1 Convolution is used to process the data to obtain features of the same scale. Initial features The initial features were obtained by using DINOv2 to extract features from the input image, which is a dataset of foggy images; basic low-frequency features were extracted. Perform bilinear upsampling and then Convolution adjustment yields reconstructed base layer features Reconstructing the characteristics of the base layer The calculation formula is ,in, Bilinear upsampling; calculation of features at the same scale With reconstruction of basic layer features The residuals between them yield detailed features. , For detailed features.

[0128] In an exemplary embodiment, the feature extraction module 502 is further configured to extract features based on initial features. Initial features are generated using two consecutive convolutions. The corresponding space-channel adaptive weight matrix is , The sigmoid function is used to adjust the spatial-channel adaptive weight matrix. With reconstruction of basic layer features Element-wise multiplication yields the enhanced basic features. , The symbol for element-wise multiplication.

[0129] In one exemplary embodiment, the construction module 504 is further configured to enhance the basic features. With detailed features Perform channel-dimensional concatenation to obtain the fused features. ; fusion features After performing convolution and normalization, the refined features are obtained. .

[0130] In an exemplary embodiment, the feature extraction module 502 is further configured to refine the features. Global average pooling and convolution are performed to obtain refined features. The channel importance weight is And assign channel importance weights Characteristics acting on refined products The contrast enhancement features obtained are ;in, Based on the characteristics of the refined The result is obtained by global average pooling, and the calculation formula is as follows: Enhance contrast and strengthen features through detailed processing. The high-frequency detail information is used to obtain detail enhancement features. ;Detail-enhanced features and initial features Determine the output characteristics of the defogging layer as follows .

[0131] In an exemplary embodiment, the feature extraction module 502 is further configured to output features to the dehazing layer. Convolution and nonlinear transformation processes are performed to obtain the output features of the dehazing layer. The importance weights of each channel are: and output features for the dehazing layer. Convolution and normalization are performed to obtain the output features of the dehazing layer. The spatial enhancement coefficient is In the formula, The function is a sigmoid function; a learnable gating network is used to fuse the output features of the dehazing layer. Importance weights of each channel and spatial enhancement coefficient The fusion enhancement coefficient is obtained as ,in, , From the spatial enhancement coefficient Obtained by extending along the channel dimension to the original number of channels; based on the fusion enhancement coefficient. and the output characteristics of the defogging layer The enhanced features are For the enhanced features Compressed projection processing is performed to obtain compact semantic features corresponding to the input image. .

[0132] In one exemplary embodiment, the fog-day cross-view geolocation device further includes:

[0133] A pre-training module is used to acquire a training dataset; the training dataset includes query images and a library image set, which are cross-view images; a cross-view localization model is constructed based on a pre-defined cross-view localization network and the training dataset; the cross-view localization network includes an adaptive feature dehazing layer and a dual-path feature enhancement layer; compact semantic features corresponding to the query images are determined based on the adaptive feature dehazing layer and the dual-path feature enhancement layer. Compact semantic features corresponding to each image in the image library set. And the compact semantic features of the positive samples corresponding to the query image. Based on the compact semantic features corresponding to the query image Compact semantic features corresponding to each image in the image library set. And the compact semantic features of the positive samples corresponding to the query image. The loss function is determined as follows: In the formula, For the first image in the image library collection Compact semantic features corresponding to images in the image library The temperature parameter is used; the cross-view localization network is optimized using the loss function to obtain the cross-view localization model.

[0134] The modules in the aforementioned fog-weather cross-view geolocation device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0135] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0136] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0137] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0138] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0139] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). 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.

[0140] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0141] The above embodiments are merely illustrative of several implementation methods of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A cross-view geolocation method in foggy weather, characterized in that, The method includes: S1, take the image to be located and the image library set as input images, input them into the pre-trained cross-view localization model, and determine the compact semantic features corresponding to each image in the image library set and the image to be located; the cross-view localization model includes an adaptive feature dehazing layer and a dual-path feature enhancement layer; The cross-view localization model processes the input image as follows: Based on the adaptive feature dehazing layer, the initial features corresponding to the input image are decomposed at multiple scales and subjected to adaptive filtering to obtain enhanced basic features. These enhanced basic features are then subjected to guided detail fusion to obtain refined features. and the refined features Contrast enhancement and detail enhancement processes are performed to obtain the output features of the dehazing layer. ; Based on the dual-path feature enhancement layer, the dehazing layer outputs features. Channel path enhancement and spatial adaptive enhancement are performed to obtain enhanced features, and the enhanced features are then compressed and projected to obtain compact semantic features corresponding to the input image. ; S2, based on the image to be located and the compact semantic features corresponding to each image in the image library, determine the target matching image corresponding to the image to be located in the image library.

2. The method according to claim 1, characterized in that, The process of performing multi-scale decomposition of the initial features corresponding to the input image based on the adaptive feature dehazing layer includes: Initial features corresponding to the input image Using a step size of 2 Convolution is used to process the data and obtain the basic low-frequency features. and the initial features Using a step size of 1 Convolution is used to process the data to obtain features of the same scale. The initial feature The initial features of the input image are obtained by using DINOv2 to extract features from the input image, wherein the input image is a foggy day image dataset; Regarding the basic low-frequency characteristics Perform bilinear upsampling and then Convolution adjustment yields reconstructed base layer features The reconstructed base layer features The calculation formula is ,in, Bilinear upsampling; Calculate the same-scale features With the reconstructed base layer features The residuals between them yield detailed features. , For detailed features.

3. The method according to claim 2, characterized in that, The adaptive filtering process includes: Based on the initial features The initial features are generated using two consecutive convolutions. The corresponding space-channel adaptive weight matrix is , For the Sigmoid function; Space-channel adaptive weight matrix With the reconstructed base layer features Element-wise multiplication yields the enhanced basic features. , The symbol for element-wise multiplication.

4. The method according to claim 3, characterized in that, The enhanced basic features are then subjected to guided detail fusion to obtain refined features. ,include: Enhanced basic features With detailed features Perform channel-dimensional concatenation to obtain the fused features. ; Fusion features After performing convolution and normalization, the refined features are obtained. .

5. The method according to claim 4, characterized in that, The refined features Contrast enhancement and detail enhancement processes are performed to obtain the output features of the dehazing layer. ,include: The refined features Global average pooling and convolution are performed to obtain the refined features. The channel importance weight is And assign channel importance weights The characteristics acting on the refined product The contrast enhancement features obtained are ;in, From the refined features The result is obtained by global average pooling, and the calculation formula is as follows: ; Enhance contrast and strengthen features through detailed enhancement processing. The high-frequency detail information is used to obtain detail enhancement features. ; Based on detail enhancement features and the initial features Determine the output characteristics of the defogging layer as follows .

6. The method according to claim 1, characterized in that, The dual-path feature enhancement layer is used to output features to the dehazing layer. Channel path enhancement and spatial adaptive enhancement are performed to obtain enhanced features, and the enhanced features are then compressed and projected to obtain compact semantic features corresponding to the input image. ,include: Output characteristics of the defogging layer Convolution and nonlinear transformation processes are performed to obtain the output features of the dehazing layer. The importance weights of each channel are: and output features to the defogging layer. Convolution and normalization are performed to obtain the output features of the dehazing layer. The spatial enhancement coefficient is In the formula, For the Sigmoid function; Employing a learnable gating network, the output features of the dehazing layer are fused. Importance weights of each channel and spatial enhancement coefficient The fusion enhancement coefficient is obtained as ,in, , From the spatial enhancement coefficient This is obtained by expanding along the channel dimension to the original number of channels; Based on fusion enhancement coefficient and the output characteristics of the defogging layer The enhanced features are For the enhanced features Compressed projection processing is performed to obtain compact semantic features corresponding to the input image. .

7. The method according to claim 1, characterized in that, The pre-training process of the cross-view localization model includes: Obtain the training dataset; the training dataset includes query images and a library image set, and the query images and the library image set are cross-view images; A cross-view localization model is constructed based on a pre-defined cross-view localization network and the training dataset; the cross-view localization network includes an adaptive feature dehazing layer and a dual-path feature enhancement layer; Based on the adaptive feature dehazing layer and dual-path feature enhancement layer, the compact semantic features corresponding to the query image are determined. Compact semantic features corresponding to each image in the image library set. And the compact semantic features of the positive samples corresponding to the query image. ; Based on the compact semantic features corresponding to the query image Compact semantic features corresponding to each image in the image library set. And the compact semantic features of the positive samples corresponding to the query image. The loss function is determined as follows: In the formula, For the first image in the image library collection Compact semantic features corresponding to images in the image library For temperature parameters; The cross-view localization network is optimized using the loss function to obtain the cross-view localization model.

8. A cross-view geolocation device for foggy weather, characterized in that, The device includes: The feature extraction module is used to take the image to be located and the image set in the image library as input images, input them into the pre-trained cross-view localization model, and determine the compact semantic features corresponding to the image to be located and each image in the image library; the cross-view localization model includes an adaptive feature dehazing layer and a dual-path feature enhancement layer; The cross-view localization model processes the input image as follows: Based on the adaptive feature dehazing layer, the initial features corresponding to the input image are decomposed at multiple scales and subjected to adaptive filtering to obtain enhanced basic features. These enhanced basic features are then subjected to guided detail fusion to obtain refined features. and the refined features Contrast enhancement and detail enhancement processes are performed to obtain the output features of the dehazing layer. ; Based on the dual-path feature enhancement layer, the dehazing layer outputs features. Channel path enhancement and spatial adaptive enhancement are performed to obtain enhanced features, and the enhanced features are then compressed and projected to obtain compact semantic features corresponding to the input image. ; The localization module is used to determine a target matching image corresponding to the image to be located in the image library based on the image to be located and the compact semantic features corresponding to each image in the image library.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images