Adaptive Pooling Method and System Based on Feature Distribution

Abstract

The present invention discloses an adaptive pooling method and system based on feature distribution. The method includes the following steps: Step 1, produce an image data set; Step 2, construct a cross-view geolocation network model with an adaptive pooling strategy based on feature distribution, and train it with the data set in Step 1; Step 3, the trained cross-view geolocation network model receives the drone or satellite view image to be located, and matches the corresponding satellite or drone image. The adaptive pooling method and system based on feature distribution of the present invention aims to guide the network to capture significant target features while reducing the sensitivity to context features, so as to achieve accurate cross-view geolocation of drones.

Description

Technical Field

[0001] The present invention belongs to the field of UAV geolocation based on intelligent vision, and relates to a feature compression strategy. Specifically, it is an adaptive pooling method and system based on feature distribution for cross-view geolocation tasks, aiming to guide the network to capture significant target features while reducing the sensitivity to context features to achieve accurate UAV cross-view geolocation. Background Art

[0002] Cross-view geolocation aims to match images of the same geographical location obtained by different platforms (such as UAVs and satellites). Most deep learning-based solutions generally include three stages: feature extraction, feature fusion, and matching. In the feature extraction stage, a backbone network represented by ResNet-50 is used to extract high-level features of the input image. The feature fusion stage uses a feature compression strategy to convert high-dimensional features into one-dimensional visual descriptors to reduce the redundant information of high-dimensional features and the network calculation cost. In the matching stage, the generated visual descriptors corresponding to cross-view images are mapped to a shared feature space, and it is constrained that features with the same geographical label have similar distributions.

[0003] Pooling is the main method of feature compression, and the most typical pooling strategies include max pooling and average pooling. It can effectively compress high-level features into one-dimensional visual descriptors with fewer or no learnable parameters, which is equivalent to the most concise feature fusion method. Existing cross-view geolocation solutions mainly focus on context information mining and visual descriptor matching, while ignoring the impact of feature compression on the results. Especially for multi-granularity features, the existing technology uses a single pooling layer to compress all features, lacking consideration of the differences in feature content. In addition, although generalized average pooling can achieve the transformation of pooling types by adjusting parameters, this adaptive adjustment method is not stable enough. Therefore, aiming at the deficiencies of the existing technology, the present invention proposes an adaptive pooling method and system based on feature distribution, aiming to introduce distance-dependent pooling initial parameters on the basis of the traditional generalized average pooling benchmark to achieve more accurate feature compression for multi-granularity features captured by the cross-view geolocation algorithm. The adaptive pooling strategy based on feature distribution proposed by the present invention can adaptively adjust the pooling type according to the spatial structure differences of high-dimensional features corresponding to cross-views, guide the cross-view geolocation network model to better retain discriminative features in the feature compression stage, thereby improving the performance of the network model. Summary of the Invention

[0004] In view of the fact that existing cross-view geolocation technologies usually use a single pooling layer to compress all features when dealing with multi-granularity features, lacking consideration of the differences in the spatial distribution of content, the present invention designs an adaptive pooling strategy based on feature distribution to achieve more refined feature compression for multi-granularity features. In view of the content differences of multi-granularity features, the present invention designs a distance-dependent pooling parameter initialization strategy to predefine different pooling types. Applying the adaptive pooling strategy based on feature distribution proposed by the present invention to existing cross-view geolocation algorithms can effectively improve the accuracy and stability of the algorithms.

[0005] The present invention adopts the following technical solutions:

[0006] An adaptive pooling strategy based on feature distribution applicable to cross-view geolocation tasks, the specific steps are as follows:

[0007] Step 1: Make an image dataset;

[0008] Step 2: Construct a cross-view geolocation network model with an adaptive pooling strategy based on feature distribution, and train it with the dataset in Step 1;

[0009] Step 3: The trained cross-view geolocation network model receives the drone or satellite view image to be located and matches the corresponding satellite or drone image.

[0010] Preferably, in Step 1, an image dataset required for network training in Step 2 is established. Specifically as follows:

[0011] Download the publicly available dataset University-1652 dataset. The original University-1652 dataset is a drone-based geolocation dataset, which contains data from three platforms, namely, pictures of 1652 university buildings from the perspectives of drones, satellites, and the ground. This dataset can be extended to two tasks, namely, object localization from the drone perspective and drone navigation. The training set includes 701 buildings of 33 universities, and the test set includes the other 951 buildings of the remaining 39 universities. There are no overlapping universities in the training and test sets. The drone and satellite perspectives in this data are used for drone view object localization. There are 37,85 drone view images in the query set, 701 truly matching satellite view images and 250 satellite view interference items in the library, and for each drone view image, there is only one truly matching satellite view image. The goal of the drone geolocation task is to find the corresponding satellite view image based on the drone view image, so as to determine the geographical location of the drone.

[0012] Preferably, Step 2 is specifically as follows:

[0013] Step 2.1: Construct the feature extraction module and feature partitioning module of the cross-perspective geolocation benchmark network model

[0014] The benchmark network model adopted in the present invention includes a drone branch and a satellite image branch, and the two branches adopt a learning method with shared weights. Each branch consists of four parts: feature extraction, feature partitioning processing, feature compression, and classification supervision. The input is a given image X, and according to the perspective of the image, it is selected to enter the drone branch or the satellite image branch; the given image X is subjected to feature extraction through the Resnet-50 backbone network to obtain the high-dimensional features corresponding to the input image:

[0015] f i =F Resnet (X i ), i ∈ {D, S} (1)

[0016] Among them, f i represents the image features extracted by ResNet-50, the subscript i indicates that the features come from the drone perspective (D) or the satellite perspective (S), and F Resnet represents the feature extraction module.

[0017] Then, the existing circular partitioning strategy is adopted to partition the obtained image features f i so as to further divide the high-dimensional features extracted by the backbone network into multi-granularity features:

[0018] P n =SP(f i ; N), n = 1, …, N (2)

[0019] Among them, f i is the extracted image feature, P n represents the multi-granularity features after being segmented by the circular partitioning strategy, and N represents the number of partitioned regions.

[0020] Step 2.2: Design an adaptive pooling module based on feature distribution for the multi-granularity features calculated for the partitioning strategy to adaptively compress the multi-granularity features.

[0021] In this step, the multi-granularity features partitioned by the cross-perspective geolocation network model constructed in Step 2.1 are defined as the input, and an adaptive pooling module based on feature distribution is used for feature compression. The adaptive pooling module based on feature distribution consists of a generalized average pooling layer and a pooling parameter initialization strategy dependent on distance.

[0022] Generalized average pooling operation: Generalized average pooling is a pooling method that lies between average pooling and max pooling. It realizes the smoothing pooling operation of features through the adjustment of parameter p (a vector with a shape of 1×1×C). When p→∞, generalized average pooling tends to max pooling; when p = 1, generalized average pooling becomes average pooling:

[0023] Specifically, for the multi-granularity feature P divided in step 2.1 n , there are C channels that use generalized average pooling for pooling, which can be expressed as:

[0024]

[0025] The two special cases are as follows:

[0026]

[0027] Among them, represents the feature of the c-th channel of the multi-granularity feature obtained by the division in step 2.1, represents the feature of the c-th channel of the output feature map processed by the pooling module, represents the parameter p for generalized average pooling, n the component in the c-th channel, and their relationship with the total feature map P n of the n-th divided block, A n and its parameter p n can be expressed as:

[0028]

[0029]

[0030] Among them, represents the real number field, and 1×1×C represents the dimension of the pooling parameter.

[0031] At the same time, the parameter p N of all blocks is defined as:

[0032]

[0033] Distance-dependent pooling parameter initialization strategy: By comparing the drone view and the satellite view, it can be observed that the spatial distributions of the content in the two perspectives are similar: discriminative features (geographical targets) are usually concentrated in the center of the view, while environmental features are distributed in the peripheral areas of the view, which may contain false targets. Specifically:

[0034] 1. Discriminative features:

[0035] In drone and satellite images, important discriminative features (such as buildings, roads, etc.) often lie in the central region of the image. This is because in the vast majority of scenarios, placing key objects at the center of the picture serves to emphasize the key objects. For example, in images taken by drones, important features such as target buildings and bridges are placed in the middle region of the image.

[0036] 2. Environmental features:

[0037] Environmental features usually appear in the peripheral regions of the image. These features include natural landscapes (such as trees, rivers) and other background information. These features can provide context information, which helps in the identification of the main target. However, due to the diversity and complexity of environmental features, they may contain many irrelevant or pseudo-targets, which to a certain extent will interfere with the processing and analysis of the image.

[0038] Based on the above characteristics, the present invention aims to retain discriminative features, make full use of context features while weakening the interference of the network model to pseudo-targets, and designs a distance-dependent pooling parameter initialization strategy: that is, setting larger pooling parameters for central features to guide the pooling type towards max pooling, so as to utilize prominent discriminative features to generate feature descriptors; while setting pooling parameters tending to 1 for peripheral context features to guide the pooling type towards average pooling, so as to comprehensively utilize context features to generate feature descriptors. To further simplify the process of obtaining the values of pooling initialization parameters, the present invention designs the function Γ(n) to initialize the parameter p N , where Γ(n) is a monotonically increasing function that increases with the increase of n.

[0039] p N ={p n |p n =Γ(n), n∈(0,N]}(9)

[0040] where Γ(1)=1≤Γ(n)∈[1,+∞]. In practice, Γ(n) can be expressed as the following function:

[0041]

[0042] Construct an adaptive pooling module based on feature distribution: The number of multi-granularity features obtained by the feature partitioning module in step 2.1 determines the number of generalized pooling layers included in the adaptive pooling module based on feature distribution proposed by the present invention. Further, a set of pooling initial parameters is generated using the monotonic function designed by formula (10) and are respectively used to initialize each generalized pooling layer in the adaptive pooling module, completing the construction of the adaptive pooling module based on feature distribution. Adopting the adaptive pooling strategy based on feature distribution proposed by the present invention as the pooling module of the cross-view geolocation network model, the above process can be expressed as:

[0043] A n = D 2 GeM(P n )(11)

[0044] Wherein, D 2 GeM represents the operation of a generalized mean pooling module with distance-based parameter adaptation.

[0045] Step 2.3: After passing through the adaptive pooling module D 2 GeM, initialize the pooling parameters based on distance, perform parameter adaptive pooling on each divided feature block, and splice them to obtain the output feature map A processed by the pooling module n . Then, through the classification module CF, obtain the output score O of the classifier module n . Finally, continuously optimize the model parameters using the cross-entropy loss function to make the drone image and the satellite image match precisely. Use the one-dimensional prediction vector to match and retrieve images from different perspectives, and then achieve cross-perspective geolocation.

[0046] The whole process can be expressed as:

[0047]

[0048] Wherein, f i represents the image features extracted by ResNet-50, and the subscript i indicates that the features come from the drone perspective (D) or the satellite perspective (S). P n represents the fine-grained features obtained after division, A n represents the one-dimensional feature vector of the output feature map processed by the pooling module. O n is the output score of the classifier module. SP represents the feature division module, D 2 GeM represents the adaptive pooling module based on feature distribution, and CF represents the classifier prediction module.

[0049] Preferably, step 3 is specifically as follows: Take the drone (or satellite) perspective image X to be located as the input data and input it into the network model. After the input data passes through feature extraction, feature division, and the pooling module in sequence, in the classifier prediction module, remove the classification layer, compare the obtained features with the features obtained from the input of the existing satellite (or drone) image, and use the Euclidean distance to measure the similarity of the two perspective images to obtain the satellite (or drone) image that matches the drone (or satellite) perspective image.

[0050] The present invention also discloses an adaptive pooling system based on feature distribution for performing the above method, including the following modules:

[0051] Dataset production module: Produce an image dataset;

[0052] Training module: Construct a cross-view geolocation network model with an adaptive pooling strategy based on feature distribution, and train it with the produced dataset;

[0053] Matching module: The trained cross-view geolocation network model receives the drone or satellite view image to be located, and matches the corresponding satellite or drone image.

[0054] The adaptive pooling method and system based on feature distribution proposed by the present invention can pre-define the pooling type according to the content differences of multi-granularity features, guide the network model to retain significant features while smoothing environmental information during the feature compression stage, weaken the network model's excessive attention to environmental features, so as to achieve more refined feature compression and improve the algorithm performance. The present invention has strong flexibility and versatility and can be applied to various cross-view geolocation algorithms. Description of the Drawings

[0055] Figure 1 is a schematic flowchart of an adaptive pooling method based on feature distribution according to a preferred embodiment of the present invention.

[0056] Figure 2 is a flowchart of a cross-view geolocation network introducing an adaptive pooling strategy based on feature distribution according to a preferred embodiment of the present invention.

[0057] Figure 3 is a schematic diagram of Generalized Mean Pooling (GeM Pooling).

[0058] Figure 4 is to add D 2 - A flowchart of a cross-view geolocation network with a GeM pooling strategy.

[0059] Figure 5 is a comparison graph of the function Γ(n) for initializing pooling parameters.

[0060] Figure 6 is a block diagram of an adaptive pooling system based on feature distribution according to a preferred embodiment of the present invention. Detailed Embodiments

[0061] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0062] The present invention first makes the following definitions and explanations:

[0063] X i: The given input image, where the subscript i indicates the source of the image, which can be the drone view (D) or the satellite view (S);

[0064] f i : Represents the image features extracted by ResNet-50;

[0065] F Resnet : Represents the feature extraction module;

[0066] SP: Represents the feature partitioning module;

[0067] CF: Represents the classifier prediction module;

[0068] D 2 GeM: Generalized mean pooling module with distance-based parameter adaptation;

[0069] p N : Used to initialize the parameters of multiple generalized mean pooling layers;

[0070] Γ(n): Function used to initialize the parameter p N and is a monotonically increasing function that increases as n increases;

[0071] This embodiment proposes a distance-based parameter adaptive generalized pooling method applicable to cross-view geolocation tasks, and its basic steps are as follows:

[0072] Step 1: Establish the image dataset required for network training in Step 2.

[0073] Download the public dataset University-1652 dataset. The original University-1652 dataset is a drone-based geolocation dataset that contains data from three platforms, namely pictures of 1652 university buildings from the drone, satellite, and ground perspectives. This dataset can be extended to two tasks, namely drone view target localization and drone navigation. The training set includes 701 buildings from 33 universities, and the test set includes the other 951 buildings from the remaining 39 universities. There are no overlapping universities in the training and test sets. Take the drone and satellite views in this data for drone view target localization. There are 37,85 drone view images in the query set, 701 real matching satellite view images and 250 satellite view distractors in the database, and for each drone view image, there is only one real matching satellite view image. The goal of the drone geolocation task is to find the corresponding satellite view image based on the drone view image to determine the geographical location of the drone.

[0074] Step 2: Construct a cross-view geolocation network model based on the adaptive pooling strategy of feature distribution, specifically:

[0075] Step 2.1: Construct the feature extraction module and feature partitioning module of the cross-view geolocation benchmark network model

[0076] The benchmark network model adopted in this embodiment includes a drone branch and a satellite image branch, and the two branches adopt a learning method with shared weights. Each branch consists of four parts: feature extraction, feature partitioning processing, feature compression, and classification supervision. The input is a given image X. According to the perspective of the image, it is selected to enter the drone branch or the satellite image branch; for the given image X, it is first scaled to a pixel size of 512×512, and feature extraction is performed through the Resnet-50 backbone network to obtain the high-dimensional features corresponding to the input image:

[0077] f i =F Resnet (X i ), i ∈ {D, S} (13)

[0078] Among them, f i represents the image features extracted by ResNet-50. The subscript i indicates that the features come from the drone perspective (D) or the satellite perspective (S), and F Resnet represents the feature extraction module.

[0079] Then, the existing circular partitioning strategy is used to partition the obtained image features f i to further divide the high-dimensional features extracted by the backbone network into multi-granularity features:

[0080] P n =SP(f; N), n = 1, …, N (14)

[0081] Among them, P n represents the multi-granularity features after being segmented by the circular partitioning strategy, and N represents the number of partitioned regions.

[0082] Step 2.2: Design an adaptive pooling module based on feature distribution for multi-granularity features to adaptively compress the multi-granularity features.

[0083] In this step, the multi-granularity features partitioned by the cross-view geolocation model constructed in Step 2.1 are defined as the input, and the proposed adaptive pooling module based on feature distribution is used for feature compression. The adaptive pooling module based on feature distribution consists of a generalized average pooling layer and a distance-dependent pooling parameter initialization strategy.

[0084] Generalized average pooling operation:

[0085] Generalized average pooling is a pooling method that lies between average pooling and max pooling. It achieves smooth pooling operation on features through the adjustment of parameter p (a vector with a shape of 1×1×C). When p→∞, generalized average pooling tends to max pooling; when p = 1, generalized average pooling becomes average pooling:

[0086] Specifically, for the multi-granularity feature P divided in step 2.1 n , there are C channels using generalized average pooling for pooling, which can be expressed as:

[0087]

[0088] The two special cases are as follows:

[0089]

[0090] Among them, represents the feature of the c-th channel of the multi-granularity feature obtained by the division in step 2.1, represents the feature of the c-th channel of the output feature map processed by the pooling module, represents the parameter p for generalized average pooling, n the component of which in the c-th channel, and their relationship with P n , A n and its parameter p n in the n-th block of the division can be expressed as:

[0091]

[0092] At the same time, the parameter p N of all blocks is defined as:

[0093]

[0094] Pooling parameter initialization strategy dependent on distance: By comparing the drone view and the satellite view, it can be observed that the spatial distributions of the contents in the two perspectives are similar: discriminative features (geographical targets) are usually concentrated in the center of the view, while environmental features are distributed in the peripheral areas of the view, which may contain false targets. Specifically:

[0095] 1. Discriminative features:

[0096] In drone and satellite images, important discriminative features (such as buildings, roads, etc.) tend to be located in the central area of the image. This is because placing key objects in the center of the picture has the effect of emphasizing key objects in the vast majority of scenarios. For example, in the images taken by drones, important features such as target buildings and bridges will be placed in the middle area of the image.

[0097] 2. Environmental characteristics:

[0098] Environmental characteristics usually appear in the peripheral areas of images. These characteristics include natural landscapes (such as trees, rivers) and other background information. These characteristics can provide context information, which helps in the identification of the main target. However, due to the diversity and complexity of environmental characteristics, they may contain many irrelevant or pseudo-targets, which to a certain extent will interfere with the processing and analysis of images.

[0099] Based on the above characteristics, this embodiment aims to retain discriminative features, make full use of context features while weakening the interference of the network model on pseudo-targets, and designs a distance-dependent pooling parameter initialization strategy: that is, set larger pooling parameters for central features to guide the pooling type towards max pooling, so as to use prominent discriminative features to generate feature descriptors; while set pooling parameters tending to 1 for peripheral context features to guide the pooling type towards average pooling, so as to comprehensively use context features to generate feature descriptors. To further simplify the process of obtaining the values of pooling initialization parameters, this embodiment designs the function Γ(n) to initialize the parameter p N , where Γ(n) is a monotonically increasing function that increases with the increase of n.

[0100] p N ={p n |p n =Γ(n), n ∈ (0, N]}(21)

[0101] where Γ(1) = 1 ≤ Γ(n) ∈ [1, +∞]. In practice, Γ(n) can be expressed as the following function (see Figure 5 ):

[0102]

[0103] Construct an adaptive pooling module based on feature distribution: The number of multi-granularity features obtained by the feature division module in step 2.1 determines the number of generalized pooling layers included in the adaptive pooling module based on feature distribution proposed by the present invention. Further, use the monotonic function proposed in formula (22) to generate a set of pooling initial parameters, and respectively use them to initialize each generalized pooling layer in the adaptive pooling module to complete the construction of the adaptive pooling module based on feature distribution. Using the adaptive pooling strategy based on feature distribution proposed in this embodiment as the pooling module of the cross-view geolocation network model, the above process can be expressed as:

[0104] A n =D 2 GeM(P n )(23)

[0105] where, D 2GeM represents an adaptive pooling module based on feature distribution.

[0106] Step 2.3: Construct a cross-view geolocation network model that incorporates an adaptive pooling strategy based on feature distribution and train it using a dataset.

[0107] The specific structure of the complete network is as Figure 2 shown. The network model includes a drone branch and a satellite image branch, and the two branches adopt a shared-weight learning method. Each branch consists of four parts: feature extraction, feature partitioning processing, and classification supervision. The initial input of the network is a given image X, whose size is adjusted to a pixel size of 512×512. According to the perspective of the image, it is selected to enter either the drone branch or the satellite image branch; for the given image X, it first passes through the feature extraction module F Resnet to obtain the feature map f i ; then it passes through the feature partitioning module SP, which first partitions the feature map using an annular partitioning strategy to obtain the partitioned fine-grained features P n ; subsequently, it passes through the adaptive pooling module D 2 GeM based on feature distribution, initializes the pooling parameters based on distance, performs parameter adaptive pooling on each partitioned feature block, and then concatenates them to obtain the output feature map A n after processing by the pooling module. Then it passes through the classification module CF to obtain the prediction vector O n output by the classifier module. Finally, the model parameters are continuously optimized using the cross-entropy loss function to make the drone images and satellite images match precisely. The one-dimensional prediction vector is used for image matching and retrieval under different perspectives, thereby achieving cross-view geolocation.

[0108] The whole process can be expressed as:

[0109]

[0110] Among them, f i represents the image features extracted by ResNet-50, and the subscript i indicates that the features come from the drone perspective (D) or the satellite perspective (S). P n represents the multi-grained features obtained after partitioning, A n represents the one-dimensional feature vector of the output feature map after processing by the pooling module. O n is the prediction vector output by the classifier module, whose shape is the same as the number C of geographical labels. SP represents the feature partitioning module, D 2 GeM represents the adaptive pooling module based on feature distribution, and CF represents the classifier prediction module.

[0111] Step 3: The trained neural network model receives the drone perspective image to be located and matches the corresponding satellite image for it.

[0112] First, the drone (or satellite) perspective image X to be located (adjusted to a pixel size of 512×512) is used as input data and fed into the network model. After the input data passes through the feature extraction, feature partitioning, and pooling modules in sequence, in the classifier prediction module, the classification layer is removed to obtain a prediction vector with a shape of 1×C. The obtained features are compared with the features obtained from the input of the existing satellite (or drone) images, and the Euclidean distance is used to measure the similarity of the two perspective images, so as to obtain the satellite image that matches the drone (or satellite) perspective image.

[0113] The results of using different Γ(·) functions in the distance-based parameter adaptive generalized average pooling strategy added through experiments are shown in Table 1.

[0114] Table 1 Results of using different pooling methods on the University-1652 dataset based on the benchmark model

[0115]

[0116] In the experiment, the following implementation details were followed: The input images were adjusted to 512×512 pixels, the stochastic gradient descent optimizer was used, and the size of each batchsize was 8. For the backbone network, the initial learning rate was set to 0.001, while for other layers, the initial learning rate was set to 0.01. A total of 120 epochs were carried out during the training process, and the learning rate was reduced by 0.1 times after the 40th and 80th epochs. All models were implemented using the PyTorch framework and trained on an NVIDIA GTX 3090 GPU with 24GB of memory.

[0117] As can be seen from Table 1 above, after adding the pooling strategy proposed in the present invention, it is observed that the effect of the model has been greatly improved.

[0118] As Figure 6 shown, this embodiment discloses an adaptive pooling system based on feature distribution for implementing the above method embodiment, which includes the following modules:

[0119] Dataset production module: Produce an image dataset;

[0120] Training module: Construct a cross-perspective geolocation network model based on the adaptive pooling strategy of feature distribution and train it through the produced dataset;

[0121] Matching module: The trained cross-perspective geolocation network model receives the drone or satellite perspective image to be located and matches the corresponding satellite or drone image.

[0122] For other content of this embodiment, reference may be made to the above method embodiment.

[0123] The above specific embodiments are only used to explain the present invention, rather than all embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts are within the protection scope of the present invention.

Claims

1. An adaptive pooling method based on feature distribution, characterized by It includes the following steps: Step 1: Make an image dataset; Step 2: Construct a cross-view geolocation network model based on an adaptive pooling strategy of feature distribution, and train it with the dataset in Step 1; use an adaptive pooling module based on feature distribution for feature compression, and the adaptive pooling module based on feature distribution includes generalized mean pooling and a distance-dependent pooling parameter initialization strategy; For the divided multi-granularity feature P n , C channels are pooled using generalized average pooling, expressed as: The two special cases are as follows: Among them, represents the feature of the c-th channel of the multi-granularity feature P obtained by partitioning n ; represents the feature of the c-th channel of the output feature map A processed by the pooling module n ; represents the parameter p for generalized average pooling n in the c-th channel component, which are related to P of the n-th block of the partition n , A n and its parameter p n are expressed as: Among them, represents the real number field, and 1×1×C represents the dimension of the pooling parameter; At the same time, the parameter p of all blocks N is defined as: Design the function Γ(n) to initialize the parameter p N , where Γ(n) is a monotonically increasing function that increases with the increase of n; p N = {p n | p n = Γ(n), n ∈ (0, N]}(9) Among them, Γ(1) = 1 ≤ Γ(n) ∈ [1, +∞], and Γ(n) is expressed as the following function: ... indicates that other different functions can also be used in different scenarios; Generate a set of pooling initial parameters using the monotonic function designed by formula (10), and use them to initialize each generalized pooling layer in the adaptive pooling module to complete the construction of the adaptive pooling module based on feature distribution; adopt the adaptive pooling strategy based on feature distribution as the pooling module of the cross-view geolocation network model, which is expressed as: A n = D 2 GeM(P n )(11) Among them, D 2 GeM represents the operation of a generalized mean pooling module with distance-based parameter adaptation; Step 3: The trained cross-view geolocation network model receives the drone or satellite perspective image to be located and matches the corresponding satellite or drone image.

2. The adaptive pooling method based on feature distribution according to claim 1, characterized in that: In Step 1, download the publicly available University-1652 dataset.

3. The adaptive pooling method based on feature distribution according to claim 1, characterized in that: Step 2 is specifically as follows: Step 2.1: Construct a feature extraction module and a feature partitioning module of the cross-view geolocation benchmark network model; Step 2.2: Design the adaptive pooling module based on feature distribution for the multi-granularity features calculated for the partitioning strategy to adaptively compress the multi-granularity features; Step 2.3: Introduce the adaptive pooling strategy based on feature distribution, construct a cross-view geolocation network model, and train it with the dataset in Step 1.

4. The adaptive pooling method based on feature distribution according to claim 3, wherein: Step 2.1 is specifically as follows: The benchmark network model includes a drone branch and a satellite image branch, and the two branches adopt a learning method with shared weights; each branch consists of four parts: feature extraction, feature partitioning processing, feature compression, and classification supervision; the input is a given image X, and according to the perspective of the image, it is selected to enter the drone branch or the satellite image branch; the given image X is subjected to feature extraction through a Resnet-50 backbone network to obtain the high-dimensional features corresponding to the input image: f i = F Resnet (X i ), i ∈ {D, S}(1) Among them, f i represents the image features extracted by the ResNet-50 backbone network, X i is the given input image, i indicates that the features come from the drone view D or the satellite view S, and F Resnet represents the feature extraction module; Use a circular partitioning strategy to partition the obtained image feature f i for further partitioning the high-dimensional features extracted by the backbone network into multi-granularity features: P n = SP(f i ; N), n = 1, …, N(2) Among them, SP represents the feature partitioning module; P n represents the multi-granularity features after being segmented by the circular partitioning strategy, and N represents the number of partitioned regions; Step 2.2 is specifically as follows: Use the multi-granularity features partitioned by the cross-view geolocation network model constructed in Step 2.1 as the input; Step 2.3 is specifically as follows: Initialize the pooling parameters based on the distance, perform parameter adaptive pooling on each divided feature block, and perform splicing to obtain the output feature map A processed by the pooling module n ; Through the classifier prediction module CF, obtain the output score O of the classifier prediction module n ; Finally, use the cross-entropy loss function LOSS to optimize the model parameters to match the drone image with the satellite image; Use the one-dimensional prediction vector to match and retrieve images from different perspectives to achieve cross-perspective geolocation It is expressed as: Among them, Y represents the true class label of the input image, C represents the total number of classes, and P n represents the multi-granularity features obtained after partitioning, and A n represents the one-dimensional feature of the output feature map after processing by the pooling module, and O n (c) is the output score of the classifier module for class c, Y represents the true class label of the input image, C represents the total number of classes, and D 2 GeM represents the adaptive pooling module based on feature distribution, and CF represents the classifier prediction module.

5. The adaptive pooling method based on feature distribution according to claim 4, characterized in that: Step 3 is specifically as follows: Use the drone or satellite perspective image to be located as the input data and input it into the cross-view geolocation network model. After the input data passes through feature extraction, feature partitioning, and pooling module processing in sequence, in the classifier prediction module, remove the classification layer, compare the obtained features with the features obtained by inputting the existing satellite or drone image, and use the Euclidean distance to measure the similarity of the two perspective images to obtain the satellite or drone image that matches the drone or satellite perspective image.

6. An adaptive pooling system based on feature distribution for performing the method according to any one of claims 1-5, characterized in that It includes the following modules: Dataset production module: Make an image dataset; Training module: Construct a cross-view geolocation network model with an adaptive pooling strategy based on feature distribution and train it using the prepared dataset; Matching module: The trained cross-view geolocation network model receives the drone or satellite perspective image to be located and matches the corresponding satellite or drone image.

Images