Adaptive pooling method and system based on feature distribution

Abstract

The invention discloses a feature distribution-based adaptive pooling method and system. The method comprises the following steps of 1, making an image data set; step 2, constructing a cross-view geographic positioning network model of an adaptive pooling strategy based on feature distribution, and training through the data set in the step 1; and step 3, receiving an unmanned aerial vehicle or satellite view angle image needing to be positioned by the trained cross view angle geographic positioning network model, and matching a corresponding satellite or unmanned aerial vehicle image. The invention discloses a feature distribution-based adaptive pooling method and system, and aims to guide a network to capture significant target features and reduce the sensitivity to context features so as to realize accurate unmanned aerial vehicle cross-view geographic positioning.

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, the present invention is an adaptive pooling method and system based on feature distribution suitable for cross-view geolocation tasks, which aims to guide the network to capture significant target features while reducing sensitivity to contextual features, so as to achieve accurate cross-view geolocation of UAV. Background Art

[0002] Cross-view geolocation aims to match images of the same geographic location acquired by different platforms (such as drones and satellites). Most deep learning-based solutions usually 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 and network computational cost of high-dimensional features. In the matching stage, the visual descriptors corresponding to the generated cross-view images are mapped to a shared feature space, and features with the same geographic tags are constrained to have similar distributions.

[0003] Pooling is the main method of feature compression, among which the most typical pooling strategies include maximum 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 simplest feature fusion method. Existing cross-view geo-location solutions mainly focus on contextual information mining and visual descriptor matching, while ignoring the impact of feature compression on the results. Especially for multi-granularity features, the prior art uses a single pooling layer to compress all features, lacking consideration of feature content differences. 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, in view of the shortcomings of the prior art, the present invention proposes an adaptive pooling method and system based on feature distribution, aiming to introduce distance-dependent pooling initial parameters on the traditional generalized average pooling benchmark, and realize more accurate feature compression for multi-granularity features captured by the cross-view geo-location algorithm. The adaptive pooling strategy based on feature distribution proposed in the present invention can adaptively adjust the pooling type according to the spatial structure differences of high-dimensional features corresponding to the cross-view, and guide the cross-view geo-location network model to better retain the 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 the existing cross-view geolocation technology usually uses a single pooling layer to compress all features when processing multi-granularity features, and lacks consideration of the differences in content spatial distribution, the present invention designs an adaptive pooling strategy based on feature distribution to achieve more refined feature compression of 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 in the present invention to the existing cross-view geolocation algorithm can effectively improve the accuracy and stability of the algorithm.

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

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

[0007] Step 1: Create an image dataset;

[0008] Step 2: Build a cross-view geolocation network model with an adaptive pooling strategy based on feature distribution and train it with the dataset from step 1.

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

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

[0011] Download the public dataset University-1652 dataset. The original University-1652 dataset is a drone-based geolocation dataset. It contains data from three platforms, namely, images of 1,652 university buildings from drone, satellite, and ground perspectives. This dataset can be expanded to two tasks, namely, target positioning from drone perspective 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. The drone and satellite perspectives in this data are used for drone view target positioning. There are 37,85 drone view images in the query set, 701 true matching satellite view images and 250 satellite view interference items in the library, and for each drone view image, there is only one true 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 geographic location of the drone.

[0012] Preferably, step 2 is as follows:

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

[0014] The baseline network model used in this invention 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, feature compression, and classification supervision. The input is a given image X, and according to the image's viewing angle, it is selected to enter the drone branch or the satellite image branch; the given image X is feature extracted 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). Resnet Represents the feature extraction module.

[0017] Then, the existing ring partitioning strategy is used to obtain the image feature f i Divide the high-dimensional features extracted by the backbone network into multi-granular features:

[0018] P n =SP(f i ; N), n=1,…,N(2)

[0019] Among them, f i is the extracted image feature, P n It represents the multi-granularity features after segmentation by the ring partitioning strategy, and N represents the number of divided regions.

[0020] Step 2.2: Aiming at the multi-granularity features calculated by the partitioning strategy, an adaptive pooling module based on feature distribution is designed to adaptively compress the multi-granularity features.

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

[0022] Generalized average pooling operation: Generalized average pooling is a pooling method between average pooling and maximum pooling. It achieves a smooth pooling operation on features by adjusting the parameter p (a vector of shape 1×1×C). When p→∞, generalized average pooling tends to maximum 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 pooled using generalized average pooling, which can be expressed as:

[0024]

[0025] Two special cases are as follows:

[0026]

[0027] in, Represents the features of the cth channel of the multi-granular features obtained by step 2.1. Represents the features of the cth channel of the output feature map processed by the pooling module, Indicates that it is used for The parameter p of the generalized average pooling n In the c-th channel components, they are divided into the total feature map P of the n-th block n , A n and its parameter p n The relationship can be expressed as:

[0028]

[0029]

[0030] in, represents the real number domain, and 1×1×C represents the dimension of the pooling parameter.

[0031] At the same time, the parameters p of all blocks are N Defined as:

[0032]

[0033] Distance-dependent pooling parameter initialization strategy: By comparing the drone and satellite views, we can observe that the spatial distribution of the content in the two perspectives is similar: the discriminative features (geographic targets) are usually concentrated in the center of the view, while the environmental features are distributed in the peripheral area of ​​the view, which may contain pseudo targets. Specifically:

[0034] 1. Distinguishing features:

[0035] In drone and satellite images, important discriminative features (such as buildings, roads, etc.) are often located in the center of the image. This is because placing key objects in the center of the picture has the effect of emphasizing key objects in most scenarios. For example, in images taken by drones, important features such as target buildings and bridges will be placed in the middle of the image.

[0036] 2. Environmental characteristics:

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

[0038] Based on the above characteristics, the present invention aims to retain the discriminative features, make full use of the contextual features, and weaken the interference of the network model on the pseudo-targets. A distance-dependent pooling parameter initialization strategy is designed: that is, a larger pooling parameter is set for the central feature to guide the pooling type to change to the maximum pooling, so as to use the prominent discriminative features to generate feature descriptors; and a pooling parameter approaching 1 is set for the surrounding contextual features to guide the pooling type to change to the average pooling, so as to comprehensively use the contextual features to generate feature descriptors. In order to further simplify the process of taking the value of the pooling initialization parameter, 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 n.

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

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

[0041]

[0042] Constructing an adaptive pooling module based on feature distribution: The number of multi-granular features obtained by the feature partitioning module in step 2.1 determines the number of generalized pooling layers contained in the adaptive pooling module based on feature distribution proposed in the present invention. Further, a set of pooling initial parameters is generated using the monotonic function designed by formula (10), and is used to initialize each generalized pooling layer in the adaptive pooling module, thereby completing the construction of the adaptive pooling module based on feature distribution. The adaptive pooling strategy based on feature distribution proposed in the present invention is used 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] Among them, D 2 GeM stands for generalized average pooling module operation with distance-based parameter adaptation.

[0045] Step 2.3: After the adaptive pooling module D based on feature distribution 2 GeM, based on the distance initialization pooling parameters, performs parameter adaptive pooling on each feature block after division, and splices them to obtain the output feature map A processed by the pooling module n After passing through the classification module CF, the output score O of the classifier module is obtained n Finally, the cross entropy loss function is used to continuously optimize the model parameters so that the drone image and the satellite image can be accurately matched. The one-dimensional prediction vector is used to match and retrieve images from different perspectives, thereby achieving cross-perspective geolocation.

[0046] The whole process can be described as:

[0047]

[0048] 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). n A represents the fine-grained features obtained after division. n Represents the one-dimensional feature vector of the output feature map after the pooling module is processed. n is the output score of the classifier module. 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.

[0049] Preferably, step 3 is as follows: the drone (or satellite) perspective image X to be located is passed into the network model as input data, and the input data is sequentially subjected to feature extraction, feature division, and pooling modules. In the classifier prediction module, the classification layer is removed, and the obtained features are compared with the features obtained by inputting the existing satellite (or drone) image. The similarity of the two perspective images is measured using the Euclidean distance to obtain a 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, which is used to execute the above method and includes the following modules:

[0051] Dataset creation module: create image datasets;

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

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

[0054] The feature distribution-based adaptive pooling method and system proposed in the present invention can predefine the pooling type according to the content difference of multi-granular 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, thereby achieving more refined feature compression and improving algorithm performance. The present invention has strong flexibility and versatility and can be applied to various cross-perspective geolocation algorithms. BRIEF DESCRIPTION OF THE DRAWINGS

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

[0056] Figure 2 It is a flow chart of a cross-perspective geolocation network that introduces an adaptive pooling strategy based on feature distribution in a preferred embodiment of the present invention.

[0057] Figure 3 It is a schematic diagram of generalized average pooling (GeM Pooling).

[0058] Figure 4 Yes Join D 2 - Flowchart of the cross-perspective geolocation network of the GeM pooling strategy.

[0059] Figure 5 Comparison graph of the function Γ(n) used to initialize the pooling parameters.

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

[0061] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with 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 carries out the following definitions and descriptions:

[0063] X i: Given an input image, the subscript i indicates the source of the image, which can be a drone view (D) or a 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 feature partitioning module;

[0067] CF: represents the classifier prediction module;

[0068] D 2 GeM: Generalized Average Pooling Module with Distance-Based Parameter Adaptation;

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

[0070] Γ(n): used to initialize the parameter p N The function is a monotonically increasing function that increases with n.

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

[0072] Step 1: Create 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. It contains data from three platforms, namely, images of 1,652 university buildings from drone, satellite, and ground perspectives. This dataset can be expanded to two tasks, namely, target positioning from drone perspective 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. The drone and satellite perspectives in this data are used for drone view target positioning. There are 37,85 drone view images in the query set, 701 true matching satellite view images and 250 satellite view interference items in the library, and for each drone view image, there is only one true 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 geographic 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 baseline network model used in this embodiment 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, feature compression, and classification supervision. The input is a given image X. According to the image's viewing angle, choose to enter the drone branch or the satellite image branch; for a given image X, first scale it to a pixel size of 512×512, and extract features 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). Resnet Represents the feature extraction module.

[0079] Then, the existing ring partitioning strategy is used to first obtain the image feature f i Divide the high-dimensional features extracted by the backbone network into multi-granular features:

[0080] P n =SP(f;N),n=1,…,N(14)

[0081] Among them, P n It represents the multi-granularity features after segmentation by the ring partitioning strategy, and N represents the number of divided regions.

[0082] Step 2.2: For multi-granular features, design an adaptive pooling module based on feature distribution to adaptively compress multi-granular features.

[0083] In this step, the multi-granular features divided by the cross-view geolocation model constructed in step 2.1 are defined as input, and feature compression is performed using the proposed adaptive pooling module based on feature distribution. 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 between average pooling and maximum pooling. It achieves smooth pooling of features by adjusting the parameter p (a vector of shape 1×1×C). When p→∞, generalized average pooling tends to maximum 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 pooled using generalized average pooling, which can be expressed as:

[0087]

[0088] Two special cases are as follows:

[0089]

[0090] in, Represents the features of the cth channel of the multi-granular features obtained by step 2.1. Represents the features of the cth channel of the output feature map processed by the pooling module, Indicates that it is used for The parameter p of the generalized average pooling n In the c-th channel components, they are divided into n-th block P n , A n and its parameter p n The relationship can be expressed as:

[0091]

[0092] At the same time, the parameters p of all blocks are N Defined as:

[0093]

[0094] Distance-dependent pooling parameter initialization strategy: By comparing the drone and satellite views, we can observe that the spatial distribution of the content in the two perspectives is similar: the discriminative features (geographic targets) are usually concentrated in the center of the view, while the environmental features are distributed in the peripheral area of ​​the view, which may contain pseudo targets. Specifically:

[0095] 1. Distinguishing features:

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

[0097] 2. Environmental characteristics:

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

[0099] Based on the above characteristics, this embodiment aims to retain the discriminative features, make full use of the contextual features, and weaken the interference of the network model on the pseudo-targets. A distance-dependent pooling parameter initialization strategy is designed: that is, a larger pooling parameter is set for the central feature to guide the pooling type to change to the maximum pooling, so as to use the prominent discriminative features to generate feature descriptors; and a pooling parameter approaching 1 is set for the surrounding contextual features to guide the pooling type to change to the average pooling, so as to comprehensively use the contextual features to generate feature descriptors. In order to further simplify the process of taking the value of the pooling initialization parameter, this embodiment designs the function Γ(n) to initialize the parameter p N , where Γ(n) is a monotonically increasing function that increases with 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] Constructing an adaptive pooling module based on feature distribution: The number of multi-granular 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 in the present invention. Further, a set of pooling initial parameters is generated using the monotonic function proposed in formula (22), and is used to initialize each generalized pooling layer in the adaptive pooling module, thereby completing the construction of the adaptive pooling module based on feature distribution. The adaptive pooling strategy based on feature distribution proposed in this embodiment is used 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 )(twenty three)

[0105] Among them, D 2GeM stands for feature distribution based adaptive pooling module.

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

[0107] The specific structure of the complete network is as follows Figure 2 As shown in Figure 1, the network model includes a drone branch and a satellite image branch. The two branches adopt a shared weight learning method. Each branch consists of four parts: feature extraction, feature partitioning and classification supervision. The initial input of the network is a given image X, which is resized to a pixel size of 512×512. Depending on the image's viewing angle, it is selected to enter the drone branch or the satellite image branch. Given an image X, it first passes through the feature extraction module F Resnet , get the feature map f i ; Then, after the feature partitioning module SP, the feature map is first partitioned using the ring partitioning strategy to obtain the fine-grained feature P obtained by the partitioning n ; Then pass through the adaptive pooling module D based on feature distribution 2 GeM, based on the distance initialization pooling parameters, performs parameter adaptive pooling on each feature block after division, and splices them to obtain the output feature map A processed by the pooling module n After passing through the classification module CF, the prediction vector O output by the classifier module is obtained. n Finally, the cross entropy loss function is used to continuously optimize the model parameters so that the drone image and the satellite image can be accurately matched. The one-dimensional prediction vector is used to match and retrieve images from different perspectives, thereby achieving cross-perspective geolocation.

[0108] The whole process can be described 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). n Represents the multi-granularity features obtained after division, A n Represents the one-dimensional feature vector of the output feature map after the pooling module is processed. n is the prediction vector output by the classifier module, and its shape is the same as the number of geographic tags C. 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’s view image that needs to be located and matches it with the corresponding satellite image.

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

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

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

[0115]

[0116] In the experiments, the following implementation details were followed: the input images were resized to 512×512 pixels, and a stochastic gradient descent optimizer was used with a batch size of 8. The initial learning rate was set to 0.001 for the backbone network, while it was set to 0.01 for other layers. A total of 120 epochs were conducted during training, and the learning rate was reduced by a factor of 0.1 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] It can be seen from Table 1 above that after adding the pooling strategy proposed in the present invention, it is observed that the effect of the model is greatly improved.

[0118] like Figure 6 As shown, this embodiment discloses an adaptive pooling system based on feature distribution, which is used to execute the above method embodiment, and includes the following modules:

[0119] Dataset creation module: create image datasets;

[0120] Training module: build a cross-view geolocation network model with an adaptive pooling strategy based on feature distribution and train it with the prepared dataset;

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

[0122] For other contents of this embodiment, please refer to the above method embodiment.

[0123] The above specific embodiments are only used to explain the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the protection scope of the present invention.

Claims

1. Adaptive pooling method based on feature distribution, characterized by The steps include: Step 1: Create an image dataset; Step 2: Build a cross-view geolocation network model with an adaptive pooling strategy based on feature distribution and train it with the dataset from step 1. Step 3: The trained cross-view geolocation network model receives the drone or satellite view images that need to be located and matches the corresponding satellite or drone images.

2. The adaptive pooling method based on feature distribution as claimed in claim 1, characterized in that: In step 1, download the public University-1652 dataset.

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

4. The adaptive pooling method based on feature distribution as claimed in claim 3, characterized in that: Step 2.1 is as follows: The baseline network model includes a drone branch and a satellite image branch. The two branches adopt a shared weight learning method. Each branch consists of four parts: feature extraction, feature partitioning, feature compression, and classification supervision. The input is a given image X. According to the image's viewing angle, it is selected to enter the drone branch or the satellite image branch. The given image X is feature extracted through the 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 For a given input image, i indicates that the feature comes from the drone view D or the satellite view S, F Resnet represents the feature extraction module; The obtained image feature f is divided into two parts using the ring partitioning strategy. i Divide the high-dimensional features extracted by the backbone network into multi-granular features: P n =SP(f i ;N),n=1,…,N(2) Among them, SP represents the feature partitioning module; P n It represents the multi-granularity features after segmentation by the ring partitioning strategy, and N represents the number of divided regions; Step 2.2 is as follows: taking the multi-granularity features divided by the cross-view geolocation network model constructed in step 2.1 as input, and using the feature distribution-based adaptive pooling module to perform feature compression; the feature distribution-based adaptive pooling module includes generalized average pooling and distance-dependent pooling parameter initialization strategies; For the multi-granularity feature P divided in step 2.1 n , there are C channels pooled using generalized average pooling, expressed as: Two special cases are as follows: in, Represents the features of the cth channel of the multi-granular features obtained by step 2.

1. Represents the features of the cth channel of the output feature map processed by the pooling module, Indicates that it is used for The parameter p of the generalized average pooling is n In the c-th channel components, they are divided into n-th block P n , A n and its parameter p n The relationship is expressed as: in, represents the real number domain, 1×1×C represents the dimension of the pooling parameter; At the same time, the parameters p of all blocks are N Defined as: Design function Γ(n) initialization parameter p N , where Γ(n) is a monotonically increasing function that increases with n; p N ={p n |p n =Γ(n),n∈(0,N]}(9) Among them, Γ(1)=1≤Γ(n)∈[1,+∞], Γ(n) is expressed as the following function: ...indicates that other different functions can be used in different scenarios; A set of initial pooling parameters is generated by the monotonic function designed by formula (10), and used to initialize the generalized pooling layers in the adaptive pooling module to complete the construction of the adaptive pooling module based on feature distribution. The adaptive pooling strategy based on feature distribution is used 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 generalized average pooling module operation with distance-based parameter adaptation; Step 2.3 is as follows: Initialize the pooling parameters based on the distance, perform parameter adaptive pooling on each feature block after division, and splice them to obtain the output feature map A processed by the pooling module n ; After the classifier prediction module CF, the output score O of the classifier prediction module is obtained n ; Finally, the cross entropy loss function LOSS is used to optimize the model parameters to match the drone image with the satellite image; the one-dimensional prediction vector is used to match and retrieve images from different perspectives to achieve cross-perspective geographic positioning; It is expressed as: Among them, Y represents the true category label of the input image, C represents the total number of categories, and P n Represents the multi-granularity features obtained after division, A n Represents the one-dimensional feature of the output feature map after the pooling module is processed, O n (c) is the output score of the classifier module for category c, Y represents the true category label of the input image, C represents the total number of categories, 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 as claimed in claim 4, characterized in that: Step 3 is as follows: the drone or satellite view image to be located is passed as input data into the cross-view geolocation network model. The input data is processed by feature extraction, feature partitioning, and pooling modules in sequence. In the classifier prediction module, the classification layer is removed, and the obtained features are compared with the features obtained by inputting the existing satellite or drone images. The Euclidean distance is used to measure the similarity of the two view images to obtain a satellite or drone image that matches the drone or satellite view image.

6. An adaptive pooling system based on feature distribution, for executing the method according to any one of claims 1 to 5, characterized in that Includes the following modules: Dataset creation module: create image datasets; Training module: build a cross-view geolocation network model with an adaptive pooling strategy based on feature distribution and train it with the prepared dataset; Matching module: The trained cross-view geolocation network model receives the drone or satellite view images that need to be located and matches the corresponding satellite or drone images.

Images