Enhanced global image descriptor generation device and method
By using an enhanced global image descriptor generation device, high-discriminative descriptors are generated using a cropped CNN and a multi-scale feature extraction module, which solves the accuracy problem of visual localization in complex environments and achieves higher trajectory calculation accuracy and matching success rate.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIJING INST OF TECH
- Filing Date
- 2025-12-29
- Publication Date
- 2026-04-21
AI Technical Summary
In existing technologies for multi-mode composite trajectory testing, the image matching effect of visual positioning is affected by factors such as changes in illumination and motion blur, resulting in insufficient discriminative power of descriptors and affecting the accuracy of trajectory calculation.
A pre-trained Cropped CNN is used for feature extraction, combined with a pyramid generation module, a multi-scale orientation and channel feature extraction module, and a feature enhancement and fusion module to generate high-dimensional features. Finally, descriptor generation is performed to enhance the image's discriminative power and environmental adaptability.
It significantly improves the accuracy of visual positioning and trajectory calculation in complex scenarios, and enhances the success rate and robustness of image matching.
Smart Images
Figure CN121904397A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of multi-mode composite trajectory testing technology, and particularly relates to an enhanced global image descriptor generation device and a method for operating this enhanced global image descriptor generation device. Background Technology
[0002] In multi-mode fusion trajectory testing, visual localization relies on image matching, the key to which lies in extracting highly discriminative feature descriptors from aerial image sequences. However, real-world testing environments often present challenges such as drastic lighting changes, target motion blur, and cluttered backgrounds, leading to decreased image quality and weakened target features. This results in insufficient discriminative power and poor robustness of the extracted descriptors. These issues directly impact the success rate of trajectory image matching across different viewpoints and time periods, thereby weakening the visual localization effect and ultimately reducing the accuracy of multi-mode fusion trajectory resolution.
[0003] The MixVPR model, by extracting global descriptors, provides effective feature representations for visual matching, helping to address challenges such as lighting variations and motion blur in real-world scenarios. However, its core component, Cropped CNN, has insufficient responsiveness to key semantic regions in images during feature extraction, resulting in the model's generated global descriptors failing to fully capture highly discriminative local structural information. This issue limits the model's matching accuracy in complex real-world aerial photography scenarios, thus affecting visual localization performance and the accuracy of multi-modal fusion trajectory calculation. Therefore, enhancing the feature representation capabilities of the descriptors generated by MixVPR for key regions of aerial images is crucial for improving matching performance and ultimately enhancing the overall trajectory calculation accuracy. Summary of the Invention
[0004] To overcome the shortcomings of the prior art, the technical problem to be solved by the present invention is to provide an enhanced global image descriptor generation device, which can generate image global descriptors with stronger discriminative power and environmental adaptability for aerial images, significantly improve the visual positioning accuracy in complex scenes, and thus help to improve the overall trajectory calculation accuracy.
[0005] The technical solution of the present invention is: this enhanced global image descriptor generation device includes:
[0006] The feature extraction module uses a pre-trained Croped CNN to extract features from the input image and outputs an initial feature map F.
[0007] The pyramid generation module downsamples the feature map F twice to generate a feature pyramid containing three layers of feature maps.
[0008] The multi-scale orientation feature extraction module includes an orientation feature generation unit and an orientation weight calculation unit, which extracts spatial location information from feature maps of different scales F to obtain orientation features m1; the channel feature extraction module includes a channel feature generation unit and a channel weight calculation unit, which extracts channel information from feature map F to obtain channel features m2.
[0009] The feature enhancement module uses directional feature m1 and channel feature m2 to enhance F to obtain feature map F';
[0010] The feature fusion module, which consists of L cascaded multilayer perceptron blocks, performs global context information fusion on F' and outputs high-dimensional features Z.
[0011] The descriptor generation module includes a channel projection unit, a spatial compression unit, and a normalization unit. It performs dimensionality reduction and normalization processing on the high-dimensional feature Z to generate the final global image descriptor.
[0012] This invention can generate image global descriptors with stronger discriminative power and environmental adaptability for aerial images, significantly improving the visual positioning accuracy in complex scenes, and thus helping to improve the overall trajectory calculation accuracy.
[0013] An enhanced global image descriptor generation method is also provided, which includes the following steps:
[0014] (1) Input the image to be queried, and generate an initial feature map F through the feature extraction module;
[0015] (2) Input the feature map F into the feature pyramid generation module to generate feature maps of different sizes in three dimensions, which contain feature information of different scales of feature map F;
[0016] (3) Input the feature maps of different scales in step (2) into the orientation feature extraction module to obtain the orientation feature m1;
[0017] (4) Input the feature map F into the channel feature extraction module to obtain the channel feature m2;
[0018] (5) Inject directional feature m1 and channel feature m2 into feature map F using the feature enhancement module to obtain feature map F';
[0019] (6) The feature map F' is fused globally through multiple MLP blocks to output a high-dimensional feature Z;
[0020] (7) Combine the high-dimensional feature Z input descriptor generation module, perform channel dimension projection and spatial dimension compression in sequence, and finally output the global descriptor after L2 normalization.
[0021] (8) Visual location recognition is performed using the generated descriptors, by calculating the cosine between the descriptors.
[0022] Similarity is used to achieve image matching and localization. Attached Figure Description
[0023] Figure 1 This is a flowchart illustrating the operation of an enhanced global image descriptor generation apparatus (Super-MixVPR) according to the present invention.
[0024] Figure 2 This is a schematic diagram of the Super-MixVPR according to the present invention.
[0025] Figure 3 This paper presents a performance comparison of Super-MixVPR and MixVPR (Global Image Descriptor Generation Device) based on the present invention on data_easy.
[0026] Figure 4 This is a performance comparison of Super-MixVPR and MixVPR on data_difficult according to the present invention. Detailed Implementation
[0027] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0028] To make the description of this disclosure more detailed and complete, illustrative descriptions of embodiments and specific examples of the present invention are provided below; however, these are not the only forms of implementing or utilizing the specific examples of the present invention. The embodiments cover features of multiple specific examples and methods and steps for constructing and operating these specific examples, and their order. However, other specific examples may also be used to achieve the same or equivalent functions and order of steps.
[0029] This enhanced global image descriptor generation apparatus and method includes:
[0030] The feature extraction module uses a pre-trained Croped CNN to extract features from the input image and outputs an initial feature map F.
[0031] The pyramid generation module downsamples the feature map F twice to generate a feature pyramid containing three layers of feature maps.
[0032] The multi-scale orientation feature extraction module includes an orientation feature generation unit and an orientation weight calculation unit, which extracts spatial location information from feature maps of different scales F to obtain orientation features m1; the channel feature extraction module includes a channel feature generation unit and a channel weight calculation unit, which extracts channel information from feature map F to obtain channel features m2.
[0033] The feature enhancement module uses directional feature m1 and channel feature m2 to enhance F to obtain feature map F';
[0034] The feature fusion module, which consists of L cascaded multilayer perceptron blocks, performs global context information fusion on F' and outputs high-dimensional features Z.
[0035] The descriptor generation module includes a channel projection unit, a spatial compression unit, and a normalization unit. It performs dimensionality reduction and normalization processing on the high-dimensional feature Z to generate the final global image descriptor.
[0036] This invention can generate image global descriptors with stronger discriminative power and environmental adaptability for aerial images, significantly improving the visual positioning accuracy in complex scenes, and thus helping to improve the overall trajectory calculation accuracy.
[0037] Preferably, in the feature extraction module, an initial feature map F with dimensions c×h×w is extracted from the input image i using a Cropped CNN.
[0038] Preferably, in the pyramid generation module, the feature map F is downsampled twice to obtain the feature pyramid as shown in formula (1), where each layer of the pyramid is Fi (i = 1, 2, 3) and F1 = F.
[0039] {F1,F2,F3}(1)
[0040] Preferably, in the multi-scale directional feature extraction module, the feature map Fi at each scale is subjected to global average pooling along the X-axis and Y-axis respectively to generate two directionally perceptual feature maps, where the X-axis represents the width direction and the Y-axis represents the height direction. This process is defined by formula (2):
[0041]
[0042] Where Fx is a c×h×1 feature map and Fy is a c×1×w feature map, then the shape of Fy is adjusted from c×1×w to c×h×1;
[0043] Fx and the adjusted Fy are concatenated along the channel dimension, and then a 1x1 convolution kernel is used for convolution and dimensionality reduction, as shown in Equation (3);
[0044] F inter=Conv(contact[F x ,F y (3)
[0045] The intermediate features Finter are batch normalized and activated using a non-linear activation function to generate feature map F1, which has a dimension of (c / r)×(h+w)×1, where r is the channel scaling factor. F1 is split according to the length ratio of the x and y directions of the original feature map F to obtain F1x and F1y. Then, F1x and F1y are upgraded using 1x1 convolution kernels respectively, and the final attention weights are generated by the Sigmoid activation function σ, as shown in formula (4).
[0046]
[0047] Finally, the multi-scale directional feature m1 is generated, as shown in formula (5);
[0048]
[0049] Preferably, in the channel feature extraction module, the feature map F is averaged along the channel direction and then normalized to obtain the channel feature m2, as shown in formula (6);
[0050]
[0051] Preferably, in the feature enhancement unit, the X and Y direction attention weights of the calculated m1 are multiplied element-wise with the original feature map F, and the channel m2 is multiplied one by one onto each channel of F to obtain the enhanced feature map F′, as shown in formula (7);
[0052] F'=Fgm1gm2 (7)
[0053] Preferably, in the feature mixing module, the enhanced feature map F' is input into the feature mixer, which is composed of L multilayer perceptron blocks with the same structure connected in series. Each MLP block transforms and performs residual connection on each feature through a formula, thereby fusing global context information and finally outputting high-dimensional feature Z, as shown in formula (8).
[0054] Z=FM 1...L (F') (8)
[0055] Preferably, in the descriptor generation module, Z is projected onto the channel dimension through a fully connected layer W5 to map Z from R(c×n) to R(d×n), as shown in formula (9);
[0056] Z'=W5(Z·) (9)
[0057] By projecting the fully connected layer W6 onto the spatial dimension, the features are compressed from R(d×n) to R(d×r), as shown in Equation (10):
[0058] Z”=W6(Z'·) (10)
[0059] Finally, the output is flattened and L2 normalized to obtain the final global image descriptor.
[0060] like Figure 1 As shown, an enhanced global image descriptor generation apparatus is also provided, comprising the following steps:
[0061] (1) Input the image to be queried, and generate an initial feature map F through the feature extraction module;
[0062] (2) Input the feature map F into the feature pyramid generation module to generate feature maps of different sizes in three dimensions, which contain feature information of different scales of feature map F;
[0063] (3) Input the feature maps of different scales in step (2) into the orientation feature extraction module to obtain the orientation feature m1;
[0064] (4) Input the feature map F into the channel feature extraction module to obtain the channel feature m2;
[0065] (5) Inject directional feature m1 and channel feature m2 into feature map F using the feature enhancement module to obtain feature map F';
[0066] (6) The feature map F' is fused globally through multiple MLP blocks to output a high-dimensional feature Z;
[0067] (7) Input the high-dimensional feature Z into the descriptor generation module, perform channel dimension projection and spatial dimension compression in sequence, and finally output the global descriptor after L2 normalization.
[0068] (8) Visual location recognition is performed using the generated descriptors, and image matching and localization are achieved by calculating the cosine similarity between descriptors;
[0069] Preferably, in step (1), an initial feature map F is extracted from the input image i using a Cropped CNN, with dimensions c×h×w;
[0070] Preferably, in step (2), the pyramid generation module performs two downsampling operations on the feature map F to obtain the feature pyramid as shown in formula (1), where each layer of the pyramid is Fi (i = 1, 2, 3) and F1 = F;
[0071] {F1,F2,F3}(1)
[0072] In step (3), the multi-scale directional feature extraction module performs global average pooling on the feature map Fi of each scale along the X-axis and Y-axis respectively to generate two directional perceptual feature maps. The X-axis represents the width direction and the Y-axis represents the height direction. This process is defined by formula (2).
[0073]
[0074] Where Fx is a c×h×1 feature map and Fy is a c×1×w feature map, then the shape of Fy is adjusted from c×1×w to c×h×1;
[0075] Fx and the adjusted Fy are concatenated along the channel dimension, and then a 1x1 convolution kernel is used for convolution and dimensionality reduction, as shown in Equation (3):
[0076] F inter =Conv(contact[F x ,F y (3)
[0077] The intermediate features Finter are batch normalized and activated using a non-linear activation function to generate feature map F1, which has a dimension of (c / r)×(h+w)×1, where r is the channel scaling factor. F1 is split according to the length ratio of the x and y directions of the original feature map F to obtain F1x and F1y. Then, F1x and F1y are upgraded using 1x1 convolution kernels respectively, and the final attention weights are generated by the Sigmoid activation function σ, as shown in formula (4).
[0078]
[0079] Finally, the multi-scale directional feature m1 is generated, as shown in formula (5);
[0080] In step (4), the feature map F of the channel feature extraction module is averaged along the channel direction and normalized to obtain the channel feature m2, as shown in formula (6).
[0081] In step (5), the feature enhancement unit multiplies the calculated X and Y direction attention weights of m1 with the original feature map F element by element, and multiplies the channel m2 one by one onto each channel of F to obtain the enhanced feature map F′, as shown in formula (7).
[0082] F'=Fgm1gm2 (7)
[0083] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.
[0084] The implementation process of the method of the present invention will be described in detail below, such as... Figure 2 As shown, this invention provides an enhanced global descriptor generation method (Super-MixVPR), whose core architecture includes: a feature extraction module, a pyramid generation module, a multi-scale directional feature extraction module, a channel feature extraction module, a feature mixing module, and a descriptor generation module. These modules are sequentially connected and work collaboratively. The specific construction steps are as follows:
[0085] Step 1: Remove the last global average pooling layer and fully connected layer of the Cropped CNN, and retain its convolutional layers to extract the initial feature map F from the input image i, which has dimensions c×h×w.
[0086] Step 2 is the key to improving the discriminative power of the descriptor in this invention, which injects directional and channel features into the feature map F'.
[0087] In step 2.1, the feature map F is downsampled twice in the pyramid generation module to obtain the feature pyramid as shown in equation (1). Each layer of the pyramid is Fi (i = 1, 2, 3) and F1 = F.
[0088] {F1,F2,F3}(1)
[0089] In step 2.1, the multi-scale directional feature extraction module performs global average pooling on the feature map Fi of each scale along the X-axis and Y-axis respectively to generate two directional perceptual feature maps. The X-axis represents the width direction and the Y-axis represents the height direction. This process is defined by formula (2).
[0090]
[0091] Wherein, Fx is a c×h×1 feature map, and Fy is a c×1×w feature map. Then, the shape of Fy is adjusted from c×1×w to c×h×1.
[0092] Fx and the adjusted Fy are concatenated along the channel dimension, and then a 1x1 convolution kernel is used for convolution and dimensionality reduction, as shown in Equation (3):
[0093] F inter =Conv(contact[F x ,F y (3)
[0094] Next, the obtained intermediate feature Finter is batch normalized and activated using a non-linear activation function (such as ReLU) to generate feature map F1, which has a dimension of (c / r)×(h+w)×1, where r is the channel scaling factor (set to 8 in this embodiment). F1 is split according to the length ratio of the original feature map F in the x and y directions to obtain F1x and F1y. Then, F1x and F1y are dimensionality-upgraded using 1x1 convolution kernels, and the final attention weights are generated by the Sigmoid activation function σ, as shown in formula (4):
[0095]
[0096] Finally, the multi-scale directional feature m1 is generated, as shown in equation (5);
[0097]
[0098] In step 2.2, the channel feature extraction module performs average pooling on the feature map F along the channel direction and normalizes it to obtain the channel feature m2, as shown in equation (6).
[0099]
[0100] In step 2.3, the X and Y direction attention weights of m1 are calculated and multiplied element-wise with the original feature map F, and channel m2 is multiplied one by one into each channel of F to obtain the enhanced feature map F′, as shown in formula (7):
[0101] F'=Fgm1gm2 (7)
[0102] Step 3: Feature Mixing. The enhanced feature map F' (which can be regarded as a set of two-dimensional features {Xi}) is input into the feature mixer. The mixer is composed of L MLP blocks with the same structure connected in series. Each MLP block transforms and performs residual connections on each feature through a formula, thereby fusing global context information. The final output is a high-dimensional feature Z, as shown in formula (8);
[0103] Z=FM 1...L (F') (8)
[0104] Step 4: Descriptor Generation. To generate a compact descriptor, a two-step projection method is used to reduce the dimensionality of Z;
[0105] Step 4.1 Project Z from R(c×n) to R(d×n) through the channel dimension using a fully connected layer W5, as shown in Equation (9).
[0106] Z'=W5(Z·) (9)
[0107] Step 4.2 Project the feature from R(d×n) to R(d×r) through the fully connected layer W6 in the spatial dimension, as shown in Equation (10).
[0108] Z”=W6(Z'·) (10)
[0109] Finally, in step 4.3, the output is flattened and L2 normalized to obtain the final global image descriptor.
[0110] Step 5: Performance verification.
[0111] The data for this experiment comes from the open-source dataset VPair (https: / / github.com / AerVisLoc / vpair) and Google Maps. This dataset contains satellite imagery, aerial imagery, and corresponding latitude and longitude information. Three samples, each containing 100 images, were extracted from the aerial image sequence. Based on the richness of visual feature points in the images, they were labeled as data_easy and data_difficult, with the number of feature points decreasing sequentially. Training and test sets were then created within the three datasets. Experiments were designed to verify that the image descriptors generated by the proposed Super-MixVPR model outperform the original MixVPR model. This conclusion was obtained by comparing the matching performance of the two models on the same dataset. Both experiments required extracting global descriptors from aerial and satellite images using different MixVPR models, calculating cosine similarity, and identifying the aerial-satellite image pair with the highest similarity as the matching result. If the latitude and longitude distance of the matched image pair was less than 100 meters, it was considered a correct match. A higher matching success rate indicates that the generated descriptors have stronger discriminative power and effectiveness, as shown in the experimental results. Figure 3 He Ru Figure 4In the data-easy scenario, Super-MixVPR achieved a precision of 0.43, an improvement of approximately 13% compared to MixVPR's 0.38. This improvement indicates a significant reduction in the false positive rate when returning the most relevant matches, demonstrating its stronger discriminative ability in high-level retrieval. Simultaneously, both R@9 and R@5 recall metrics improved by 0.10 (absolute value), indicating that Super-MixVPR can more consistently retrieve true matches in multi-candidate matching tasks, exhibiting superior recall robustness. Although the mean precision slightly decreased from 0.37 to 0.35, possibly due to fine-tuning the weight distribution of a few high-ranking samples when optimizing ranking confidence, the overall performance combination shows a more balanced and practical advantage. In practical applications, the simultaneous improvement in precision and recall at high candidate thresholds is often considered a key indicator of system robustness, and its value often outweighs slight fluctuations in mAP. This performance evolution demonstrates that the introduced module effectively strengthens the model's feature aggregation mechanism. By implementing dynamic context-aware weighting of spatial features, this mechanism enables the model to retain more key locational information, thereby generating global descriptors with higher recognizability and stronger viewpoint adaptability. Therefore, Super-MixVPR exhibits significantly enhanced matching and discrimination capabilities in lower-difficulty scenarios such as data_easy.
[0112] In the data-difficult scenario, Super-MixVPR's advantages are further highlighted, with performance improvements far exceeding those in the data-easy scenario. The model's accuracy improved from 0.29 to 0.44, an increase of over 50%, demonstrating a qualitative leap in its ability to identify mismatches in moderately complex environments. More importantly, Super-MixVPR achieved significant improvements in key metrics reflecting the overall retrieval capability of the system: R@9 increased from 0.30 to 0.50, an increase of approximately 67%; R@5 increased from 0.20 to 0.30, an increase of 50%; and the mAP metric, which better reflects the overall ranking quality of the system, jumped from 0.20 to 0.44, achieving a substantial improvement of 120%. These results fully demonstrate that when faced with more complex environmental factors such as changes in lighting, partial occlusion, or viewpoint shifts, the descriptors generated by Super-MixVPR can more stably capture scene structural features. By enhancing the response to key spatial locations in the feature map, the model maintains strong discriminative capabilities even under environmental interference. Therefore, in environments of varying complexity, the descriptors generated by Super-MixVPR exhibit higher precision and recall with smaller performance fluctuations, indicating stronger robustness of their feature representations. By embedding coordinate information into the channel attention mechanism, the CA module effectively enhances the model's spatial awareness. This mechanism allows the extracted global features to capture both the overall semantic information of the image and preserve local geometric structural features, thereby significantly improving the discriminative power of the MixVPR descriptors.
[0113] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention shall still fall within the protection scope of the present invention.
Claims
1. An enhanced global image descriptor generation apparatus, characterized in that: The device includes: The feature extraction module uses a cropped CNN to extract features from the input aerial image and outputs an initial feature map F. The pyramid generation module downsamples the feature map F twice to generate a feature pyramid containing three layers of feature maps. The multi-scale orientation feature extraction module includes an orientation feature generation unit and an orientation weight calculation unit, which extracts spatial location information from the feature map F at different scales to obtain orientation features m1; the channel feature extraction module includes a channel feature generation unit and a channel weight calculation unit, which extracts channel information from the feature map F to obtain channel features m2. The feature enhancement module uses directional feature m1 and channel feature m2 to enhance F to obtain feature map F'; The feature fusion module, which consists of L cascaded multilayer perceptron blocks, performs global contextual information fusion on the feature map F' and outputs high-dimensional features Z; The descriptor generation module includes a channel projection unit, a spatial compression unit, and a normalization unit. It performs dimensionality reduction and normalization processing on the high-dimensional feature Z to generate the final global image descriptor.
2. The enhanced global image descriptor generation apparatus according to claim 1, characterized in that: In the feature extraction module, an initial feature map F with dimensions c×h×w is extracted from the input image i using a cropped CNN.
3. The enhanced global image descriptor generation apparatus according to claim 2, characterized in that: In the pyramid generation module, the feature map F is downsampled twice to obtain the feature pyramid as shown in formula (1). Each layer of the pyramid is Fi (i = 1, 2, 3), where F1 = F {F1,F2,F3}(1).
4. The enhanced global image descriptor generation apparatus according to claim 3, characterized in that: In the multi-scale directional feature extraction module, the feature map Fi at each scale is subjected to global average pooling along the X-axis and Y-axis respectively, generating two directional perceptual feature maps, where the X-axis represents the width direction and the Y-axis represents the height direction. This process is defined by formula (2): Where Fx is a c×h×1 feature map and Fy is a c×1×w feature map, then the shape of Fy is adjusted from c×1×w to c×h×1; Fx and the adjusted Fy are concatenated along the channel dimension, and then a 1x1 convolution kernel is used for convolution and dimensionality reduction, as shown in Equation (3): F inter =Conv(contact[F x ,F y ]) (3) The intermediate features Finter are batch normalized and activated using a non-linear activation function to generate feature map F1, which has a dimension of (c / r)×(h+w)×1, where r is the channel scaling factor. F1 is split according to the length ratio of the x and y directions of the original feature map F to obtain F1x and F1y. Then, F1x and F1y are upgraded using 1x1 convolution kernels respectively, and the final attention weights are generated by the Sigmoid activation function σ, as shown in formula (4). Finally, the multi-scale directional feature m1 is generated, as shown in formula (5):
5. An enhanced global image descriptor generation apparatus according to claim 4, characterized in that: In the channel feature extraction module, the feature map F is averaged along the channel direction and then normalized to obtain the channel feature m2, as shown in formula (6):
6. The enhanced global image descriptor generation apparatus according to claim 5, characterized in that: In the feature enhancement unit, the calculated X and Y direction attention weights of m1 are multiplied element-wise with the original feature map F, and channel m2 is multiplied one by one onto each channel of F to obtain the enhanced feature map F′, as shown in formula (7): F'=Fgm1gm2 (7).
7. The enhanced global image descriptor generation apparatus according to claim 6, characterized in that: In the feature mixing module, the enhanced feature map F' is input into the feature mixer, which is composed of L MLP blocks with the same structure connected in series. Each MLP block transforms and performs residual connections on each feature through a formula, thereby fusing global context information and finally outputting a high-dimensional feature Z, as shown in formula (8): Z=FM 1...L (F') (8)。 8. The enhanced global image descriptor generation apparatus according to claim 7, characterized in that: In the descriptor generation module, Z is projected onto the channel dimension through a fully connected layer W5, mapping Z from R(c×n) to R(d×n), as shown in formula (9): Z'=W5(Z·) (9) By projecting the fully connected layer W6 onto the spatial dimension, the features are compressed from R(d×n) to R(d×r), as shown in Equation (10): Z”=W6(Z'·) (10) Finally, the output is flattened and L2 normalized to obtain the final global image descriptor.
9. The method of operating the enhanced global image descriptor generation apparatus according to claim 1, characterized in that: It includes the following steps: (1) Input the image to be queried, and generate an initial feature map F through the feature extraction module; (2) Input the feature map F into the feature pyramid generation module to generate feature maps of different sizes in three dimensions, which contain feature information of different scales of feature map F; (3) Input the feature maps of different scales in step (2) into the orientation feature extraction module to obtain the orientation feature m1; (4) Input the feature map F into the channel feature extraction module to obtain the channel feature m2; (5) Inject directional feature m1 and channel feature m2 into feature map F using the feature enhancement module to obtain feature map F'; (6) The feature map F' is fused globally through multiple MLP blocks to output a high-dimensional feature Z; (7) Input the high-dimensional feature Z into the descriptor generation module, perform channel dimension projection and spatial dimension compression in sequence, and finally output the global descriptor after L2 normalization. (8) Visual location recognition is performed using the generated descriptors, and image matching and localization are achieved by calculating the cosine similarity between descriptors.
10. The method of operating the enhanced global image descriptor generation apparatus according to claim 9, characterized in that: In step (1), an initial feature map F with dimensions c×h×w is extracted from the input image i using a Cropped CNN. In step (2), the pyramid generation module performs two downsampling operations on the feature map F to obtain the feature pyramid as shown in formula (1). Each layer of the pyramid is Fi (i = 1, 2, 3), F1 = F; In step (3) of {F1,F2,F3}(1), the multi-scale directional feature extraction module performs global average pooling on the feature map Fi of each scale along the X-axis and Y-axis respectively, generating two directional perceptual feature maps. The X-axis represents the width direction and the Y-axis represents the height direction. This process is defined by formula (2): Among them, F x For a c×h×1 feature map, F y For the feature map of c×1×w, then F y The shape is adjusted from c×1×w to c×h×1; Fx and the adjusted Fy are concatenated along the channel dimension, and then a 1x1 convolution kernel is used for convolution and dimensionality reduction, as shown in Equation (3): F inter =Conv(contact[F x ,F y ]) (3) The obtained intermediate features Finter are batch normalized and activated using a non-linear activation function to generate feature map F1, which has a dimension of (c / r)×(h+w)×1, where r is the channel scaling factor. F1 is split according to the length ratio of the original feature map F in the x and y directions to obtain F1x and F1y. Then, F1x and F1y are increased in dimension using 1x1 convolution kernels respectively, and the final attention weights are generated by the Sigmoid activation function σ, as shown in formula (4): Finally, the multi-scale directional feature m1 is generated, as shown in formula (5): In step (4), the feature map F of the channel feature extraction module is averaged along the channel direction and then normalized to obtain the channel feature m2, as shown in formula (6): In step (5), the feature enhancement unit multiplies the calculated X and Y direction attention weights of m1 element-wise with the original feature map F, and multiplies each channel m2 into each channel of F to obtain the enhanced feature map F′, as shown in formula (7): F'=Fgm1gm2 (7).