Remote sensing image aircraft positioning method based on all-channel space attention module

Abstract

The invention discloses a remote sensing image aircraft positioning method based on an all-channel space attention module. The method comprises the following steps: S1, obtaining a to-be-positioned remote sensing image; s2, preprocessing the to-be-positioned remote sensing image, inputting the to-be-positioned remote sensing image into a pre-trained remote sensing image aircraft positioning model, and obtaining an aircraft positioning result in the to-be-positioned remote sensing image; wherein the remote sensing image aircraft positioning model comprises a convolutional neural network, an all-channel space attention module and a target detection positioning module which are connected in sequence. The method is higher in detection precision, better in positioning accuracy, lower in computing resource demand and higher in applicability.

Description

Technical Field

[0001] The present invention relates to the technical field of target detection and positioning, and more specifically to a remote sensing image aircraft positioning method based on a full-channel spatial attention module. Background Art

[0002] Remote sensing image analysis has important application value in the fields of geographic information system (GIS), environmental monitoring, disaster assessment, agricultural management and national defense. In this field, small target detection and positioning technology is an important task in remote sensing image processing, especially in the accurate identification and positioning of small targets such as small aircraft, which has broad practical significance.

[0003] However, due to the high complexity of remote sensing images, existing small target detection and positioning methods still have problems such as low detection accuracy, poor positioning accuracy, high computing resource requirements and low applicability.

[0004] Therefore, how to provide a remote sensing image aircraft positioning method with higher detection accuracy, better positioning accuracy, lower computing resource requirements and higher applicability is a problem that technical personnel in this field urgently need to solve. Summary of the invention

[0005] In view of this, the object of the present invention is to provide a remote sensing image aircraft positioning method based on a full-channel spatial attention module.

[0006] In order to achieve the above object, the present invention adopts the following technical solution:

[0007] A remote sensing image aircraft positioning method based on a full-channel spatial attention module comprises the following steps:

[0008] S1: Acquire the remote sensing image to be located;

[0009] S2: After preprocessing the remote sensing image to be located, it is input into a pre-trained remote sensing image aircraft positioning model to obtain the aircraft positioning result in the remote sensing image to be located; wherein the remote sensing image aircraft positioning model includes a convolutional neural network, a full-channel spatial attention module and a target detection and positioning module connected in sequence.

[0010] Preferably, the convolutional neural network includes a convolution module / includes multiple convolution modules connected in sequence; wherein the convolution module includes a convolution layer, a nonlinear activation function layer, a pooling layer and a batch normalization layer connected in sequence.

[0011] Preferably, the full-channel spatial attention module includes an adaptive average pooling layer, a first 1*1 convolution layer, a first activation function layer, a second 1*1 convolution layer, an adaptive maximum pooling layer, a third 1*1 convolution layer, a second activation function layer, a fourth 1*1 convolution layer, an element-by-element addition unit, a third activation function layer, a first element-by-element multiplication unit, a channel maximum value processing unit, a channel average value processing unit, a channel splicing layer, a 7*7 convolution layer, a fourth activation function layer and a second element-by-element multiplication unit;

[0012] The input ends of the adaptive average pooling layer and the adaptive maximum pooling layer are both connected to the output end of the convolutional neural network;

[0013] The output end of the adaptive average pooling layer is connected to the input end of the element-by-element addition unit through the first 1*1 convolution layer, the first activation function layer, and the second 1*1 convolution layer in sequence;

[0014] The output end of the adaptive maximum pooling layer is connected to the input end of the element-by-element addition unit through the third 1*1 convolution layer, the second activation function layer, and the fourth 1*1 convolution layer in sequence;

[0015] The output end of the element-by-element addition unit is connected to the input end of the first element-by-element multiplication unit through the third activation function layer;

[0016] The output end of the convolutional neural network is connected to the input end of the first element-by-element multiplication unit;

[0017] The output end of the first element-by-element multiplication unit is connected to the input end of the channel maximum value processing unit and the channel average value processing unit respectively;

[0018] The output ends of the channel maximum value processing unit and the channel average value processing unit are both connected to the input end of the channel splicing layer;

[0019] The output end of the channel splicing layer is connected to the input end of the second element-by-element multiplication unit through the 7*7 convolution layer and the fourth activation function layer in sequence;

[0020] An output terminal of the first element-by-element multiplication unit is connected to an input terminal of the second element-by-element multiplication unit;

[0021] An output end of the second element-by-element multiplication unit is connected to an input end of the target detection and positioning module.

[0022] Preferably, the target detection and positioning module comprises a convolution layer, a nonlinear activation layer, a downsampling layer and an upsampling layer connected in sequence;

[0023] The input end of the convolutional layer of the target detection and positioning module is connected to the output end of the second element-by-element multiplication unit;

[0024] The output end of the upsampling layer is the output end of the remote sensing image aircraft positioning model.

[0025] Preferably, the preprocessing in S2 includes scaling, normalization and denoising.

[0026] Preferably, a pre-trained remote sensing image aircraft positioning model is obtained based on the following method:

[0027] S21: Acquire an initial remote sensing image data set; wherein the initial remote sensing image data set includes a plurality of remote sensing images including aircraft;

[0028] S22: annotating the initial remote sensing image data set; wherein the annotation content includes a bounding box of the aircraft, coordinates of the bounding box, and a category of the bounding box;

[0029] S23: preprocessing the annotated initial remote sensing image data set to obtain a final remote sensing image data set;

[0030] S24: Inputting the final remote sensing image data set into a remote sensing image aircraft positioning model to train it, thereby obtaining a pre-trained remote sensing image aircraft positioning model.

[0031] Preferably, the preprocessing in S23 includes scaling processing, normalization processing, data enhancement processing and denoising processing.

[0032] Preferably, the remote sensing image aircraft positioning model is trained based on the following total loss function:

[0033] L total =μL t +vL m +ηL s ;

[0034] Among them, L total Represents the total loss; L t , L m , L s They represent target loss, feature enhancement consistency loss, and attention sparsity loss respectively; μ, v, η represent weight coefficients.

[0035] Preferably, L t =αL box +βL cls ;

[0036]

[0037] L s =γ1L channel+γ2L spatial ;

[0038] Among them, L box represents the bounding box regression loss; L cls represents classification loss; α and β represent weight coefficients; X O Represents the original features of the convolutional neural network output; X O [i] indicates X O The i-th pixel of en represents the enhanced features output by the full-channel spatial attention module; X en [i] indicates X en The i-th pixel of O and X en The number of pixels included; L channel represents the channel attention sparsity loss; L spatial represents the spatial attention sparsity loss; γ1 and γ2 represent hyperparameters; ||X O [i]-X en [i]|| 2 Represents X O [i]-X en The square of the Euclidean norm of [i].

[0039] Preferably, L box =1-GIoU(B true , B pred );

[0040] L cls =-y c log(p c );

[0041]

[0042]

[0043] Among them, GIoU represents the overlap between the predicted bounding box and the true bounding box; B true represents the ground-truth bounding box of the annotation; B pred represents the predicted bounding box; y c One-hot encoding of the category representing the true bounding box; p c represents the classification probability of the predicted bounding box; M c Represents the channel attention map of the third activation function layer; ||M c ||1 means M c The L1-norm of M c [i, j] represents M c The pixel in the i-th row and j-th column; M s Represents the spatial attention map output by the fourth activation function layer; ||Ms ||1 means M s The L1-norm of M s [i,j] represents M s The pixel in the i-th row and j-th column; H represents M c and M s Height; W means M c and M s Width.

[0044] It can be seen from the above technical solutions that, compared with the prior art, the present invention discloses a remote sensing image aircraft positioning method based on a full-channel spatial attention module, which can achieve the following beneficial technical effects:

[0045] 1) Improved detection accuracy and positioning accuracy: By introducing a full-channel spatial attention module, the present invention significantly enhances the ability to capture small target features, making the model more superior in detecting and positioning small aircraft in remote sensing images. At the same time, the design of the total loss function effectively balances the target positioning and classification tasks, further improving the accuracy and robustness of detection. Compared with traditional methods, the present invention is particularly outstanding in detecting small targets in complex backgrounds, lighting changes or occlusions, and the positioning error is significantly reduced.

[0046] 2) Low computing resource requirements and strong applicability: The full-channel spatial attention module adopts a lightweight design, and can complete the adaptive enhancement of features through a small amount of parameter calculation, effectively reducing the computing overhead. Combined with data preprocessing and noise removal technology, the present invention maintains low resource requirements while improving model performance, so that the method of the present invention can be deployed in various computing environments. In addition, the full-channel spatial attention module provided by the present invention has strong versatility and can be flexibly embedded in the existing deep learning network architecture. It is suitable for small target detection tasks of different scales and can be widely used in remote sensing image analysis, aerial detection and other related fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0048] Figure 1 A schematic diagram of the structure of a remote sensing image aircraft positioning model provided in an embodiment of the present invention;

[0049] Figure 2 Schematic diagram of the structure of the full-channel spatial attention module provided in an embodiment of the present invention;

[0050] Figure 3 It is a schematic diagram of the structure of the target detection and positioning module provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0051] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the 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 scope of protection of the present invention.

[0052] The embodiment of the present invention discloses a remote sensing image aircraft positioning method based on a full-channel spatial attention module, comprising the following steps:

[0053] S1: Acquire the remote sensing image to be located;

[0054] S2: After preprocessing the remote sensing image to be located, the remote sensing image is input into a pre-trained remote sensing image aircraft positioning model to obtain the aircraft positioning result in the remote sensing image to be located; wherein the remote sensing image aircraft positioning model includes a convolutional neural network, a full-channel spatial attention module and a target detection and positioning module (such as Figure 1 shown).

[0055] It should be noted that the aircraft positioning result in the remote sensing image to be positioned includes the predicted bounding box of the aircraft, the coordinates of the predicted bounding box, the category of the predicted bounding box, and the classification probability of the predicted bounding box.

[0056] In a certain embodiment, the convolutional neural network includes a convolution module / includes multiple convolution modules connected in sequence; wherein the convolution module includes a convolution layer, a nonlinear activation function layer, a pooling layer and a batch normalization layer connected in sequence.

[0057] In one embodiment, if Figure 2 As shown, the full-channel spatial attention module includes an adaptive average pooling layer, a first 1*1 convolution layer, a first activation function layer, a second 1*1 convolution layer, an adaptive maximum pooling layer, a third 1*1 convolution layer, a second activation function layer, a fourth 1*1 convolution layer, an element-by-element addition unit, a third activation function layer, a first element-by-element multiplication unit, a channel maximum value processing unit, a channel average value processing unit, a channel splicing layer, a 7*7 convolution layer, a fourth activation function layer, and a second element-by-element multiplication unit;

[0058] The input ends of the adaptive average pooling layer and the adaptive maximum pooling layer are both connected to the output end of the convolutional neural network;

[0059] The output end of the adaptive average pooling layer is connected to the input end of the element-by-element addition unit through the first 1*1 convolution layer, the first activation function layer, and the second 1*1 convolution layer in sequence;

[0060] The output end of the adaptive maximum pooling layer is connected to the input end of the element-by-element addition unit through the third 1*1 convolution layer, the second activation function layer, and the fourth 1*1 convolution layer in sequence;

[0061] The output end of the element-by-element addition unit is connected to the input end of the first element-by-element multiplication unit through the third activation function layer;

[0062] The output end of the convolutional neural network is connected to the input end of the first element-by-element multiplication unit;

[0063] The output end of the first element-by-element multiplication unit is connected to the input end of the channel maximum value processing unit and the channel average value processing unit respectively;

[0064] The output ends of the channel maximum value processing unit and the channel average value processing unit are both connected to the input end of the channel splicing layer;

[0065] The output end of the channel splicing layer is connected to the input end of the second element-by-element multiplication unit through the 7*7 convolution layer and the fourth activation function layer in sequence;

[0066] An output terminal of the first element-by-element multiplication unit is connected to an input terminal of the second element-by-element multiplication unit;

[0067] An output end of the second element-by-element multiplication unit is connected to an input end of the target detection and positioning module.

[0068] It is understandable that:

[0069] For each pixel in the feature map output by the channel maximum value processing unit, the maximum value of pixels at corresponding positions in each channel feature map output by the first element-by-element multiplication unit is taken;

[0070] For each pixel in the feature map output by the channel averaging processing unit, the average value of pixels at corresponding positions in each channel feature map output by the first element-by-element multiplication unit is taken;

[0071] The first 1*1 convolutional layer and the third 1*1 convolutional layer share weights;

[0072] The first activation function layer and the second activation layer share weights;

[0073] The second 1*1 convolutional layer and the fourth 1*1 convolutional layer share weights;

[0074] The first activation function layer and the second activation function layer are both Relu functions;

[0075] The third activation function layer and the fourth activation function layer are both Sigmoid functions.

[0076] In one embodiment, if Figure 3 As shown, the target detection and positioning module includes a convolutional layer, a nonlinear activation layer, a downsampling layer and an upsampling layer connected in sequence;

[0077] The input end of the convolutional layer of the target detection and positioning module is connected to the output end of the second element-by-element multiplication unit;

[0078] The output end of the upsampling layer is the output end of the remote sensing image aircraft positioning model.

[0079] In one embodiment, the preprocessing in S2 includes scaling, normalization and denoising (the same as the scaling, normalization and denoising used in the training phase).

[0080] In one embodiment, a pre-trained remote sensing image aircraft positioning model is obtained based on the following method:

[0081] S21: Acquire an initial remote sensing image data set; wherein the initial remote sensing image data set includes a plurality of remote sensing images including aircraft;

[0082] It is understandable that: in order to ensure that the remote sensing image aircraft positioning model can handle a variety of different scenarios, thereby improving its generalization ability. The present invention collects remote sensing images from multiple sources, including but not limited to satellite images, aerial photography, and images taken by drones. These images cover different geographical areas, such as cities, rural areas, mountains, and oceans, to ensure that the model can work in a variety of environments. At the same time, the collected images also contain different weather conditions, such as clear, cloudy, rainy and snowy days, and different time periods, such as morning, noon and evening, to simulate various situations that may occur in aircraft in the real world.

[0083] S22: performing data annotation on the initial remote sensing image data set; wherein the annotation content includes a bounding box of the aircraft, coordinates of the bounding box, and a category of the bounding box;

[0084] It is understood that the present invention uses professional image annotation tools to accurately annotate the collected remote sensing images. The main task of annotation is to identify the aircraft in the image and draw a bounding box for it. The drawing of the bounding box must follow certain rules, using a rectangular box to closely fit the shape of the aircraft to ensure that every part of the aircraft is included. In addition, the annotation content also includes the coordinates of the bounding box and the category of the bounding box (the category specifically refers to whether it is an aircraft).

[0085] S23: preprocessing the annotated initial remote sensing image dataset to obtain a final remote sensing image dataset;

[0086] In one embodiment, the preprocessing in S23 includes scaling, normalization, data enhancement and denoising.

[0087] Specific:

[0088] 1) The scaling process can be expressed as: O′=Resize(O, H′, W ′ )

[0089] Wherein, O represents the remote sensing image before scaling, O′ represents the remote sensing image after scaling. H′, W′ represent the height and width of the remote sensing image after scaling;

[0090] 2) Normalization processing can be expressed as:

[0091] Among them, 0″ is the normalized remote sensing image; normalization processing means scaling the pixel value from [0, 255] to [0, 1] interval.

[0092] 3) Data enhancement processing includes random rotation, horizontal flipping, vertical flipping, random cropping and other operations;

[0093] Random rotation, horizontal flip and vertical flip operations can be expressed as: R =Rotate(O", θ) where θ represents a randomly selected rotation angle, O R Represents a rotated remote sensing image.

[0094] 4) The denoising process can be expressed as: denoise =Denoise(O R ,σ)

[0095] Where σ represents the standard deviation of the filter, O denoise The noise removal algorithm specifically applied in the present invention may be Gaussian filtering, median filtering, etc.

[0096] Through the above preprocessing steps, a series of high-quality, diverse and suitable image data for model training can be obtained. These preprocessing operations not only improve the efficiency of model training, but also enhance the accuracy and robustness of the model for aircraft positioning under different conditions.

[0097] S24: Inputting the final remote sensing image data set into a remote sensing image aircraft positioning model to train it, thereby obtaining a pre-trained remote sensing image aircraft positioning model.

[0098] Specific:

[0099] 1. Input the denoised image into the convolutional neural network to obtain the original feature X O ;

[0100] 2. The original feature X O Input to the global average pooling layer to obtain the first channel description vector X avg ;

[0101] The original feature X O Input to the global maximum pooling layer to obtain the second channel description vector X max ;

[0102] 3. Describe the first channel vector X avg Input to the first 1*1 convolution layer, the first activation function layer, and the second 1*1 convolution layer in sequence to obtain the intermediate feature map X1;

[0103] 4. The second channel description vector X max Input to the third 1*1 convolution layer, the second activation function layer, and the fourth 1*1 convolution layer in sequence to obtain the intermediate feature map X2;

[0104] 5. After adding the intermediate feature map X1 and the intermediate feature map X2 element by element, the result is input into the third activation function layer to obtain the channel attention map M c ;

[0105] 6. The channel attention map M c With the original feature X O Perform element-by-element multiplication to obtain the channel-enhanced feature X ch ;

[0106] 7. Take the channel enhanced feature X ch The maximum value of each channel in the middle feature map X3 is obtained;

[0107] Take the channel enhanced feature X ch The average value of each channel in is used to obtain the intermediate feature map X4;

[0108] It is understandable that:

[0109] Each pixel in the intermediate feature map X3 takes X ch The maximum value of the corresponding position pixel in each channel;

[0110] Each pixel of the intermediate feature map X4 is X ch The average value of the pixels at the corresponding position in each channel;

[0111] 8. The intermediate feature map X3 and the intermediate feature map X4 are concatenated in channel dimension and input into the 7*7 convolution layer and the fourth activation function layer in sequence to obtain the spatial attention map M. s ;

[0112] 9. The spatial attention map M s And the channel enhanced feature X ch Perform element-by-element multiplication to obtain enhanced feature X en .

[0113] 10. The enhancement feature X en The convolutional layer, the nonlinear activation layer, the downsampling layer, and the upsampling layer in the target detection and positioning module are sequentially input to obtain an aircraft positioning result, wherein the aircraft positioning result includes a predicted bounding box, the coordinates of the predicted bounding box, the category of the predicted bounding box, and the classification probability of the predicted bounding box (i.e., the confidence that the target in the predicted bounding box is an aircraft).

[0114] It can be understood that the present invention adopts non-maximum suppression technology to filter out non-overlapping boxes with high confidence from the predicted boundary boxes output by the model, remove redundant detection results, and obtain the aircraft positioning result.

[0115] It is understandable that the present invention can also be based on the coordinates (x min ,y min , x max ,y max ), (where (x min ,y min ) represents the coordinates of the upper left corner of the predicted bounding box; (x max ,y max ) represents the coordinates of the lower right corner of the predicted bounding box) maps the position of the predicted bounding box back to the coordinate system of the original remote sensing image. Next, the predicted bounding box, the category of the predicted bounding box, and the classification probability of the predicted bounding box (i.e., the confidence that the target in the predicted bounding box is an aircraft) are drawn on the original remote sensing image.

[0116] In one embodiment, the remote sensing image aircraft localization model is trained based on the following total loss function:

[0117] L total =μL t +vL m +ηL s ;

[0118] Among them, L total Represents the total loss; L t , L m , L sThey represent target loss, feature enhancement consistency loss, and attention sparsity loss respectively; μ, v, η represent weight coefficients.

[0119] In one embodiment, L t =αL box +βL cls ;

[0120]

[0121] L s =γ1L channel +γ2L spatial ;

[0122] Among them, L box represents the bounding box regression loss; L cls represents classification loss; α and β represent weight coefficients; X O Represents the original features of the convolutional neural network output; X O [i] indicates X O The i-th pixel of en represents the enhanced features output by the full-channel spatial attention module; X en [i] indicates X en The i-th pixel of O and X en The number of pixels included; L channel represents the channel attention sparsity loss; L spatial represents the spatial attention sparsity loss; γ1 and γ2 represent hyperparameters used to balance the channel attention sparsity loss and the spatial attention sparsity loss; ||X O [i]-X en [i]|| 2 Represents X O [i]-X en The square of the Euclidean norm of [i].

[0123] In one embodiment, L box =1-GIoU(B true , B pred );

[0124] L cls =-y c log(p c );

[0125]

[0126]

[0127] Among them, GIoU represents the overlap between the predicted bounding box and the true bounding box; Btrue represents the ground-truth bounding box of the annotation; B pred represents the predicted bounding box; y c One-hot encoding of the category representing the true bounding box; p c represents the classification probability of the predicted bounding box; M c Represents the channel attention map of the third activation function layer; ||M c ||1 means M c The L1-norm of M c [i, j] represents M c The pixel in the i-th row and j-th column; M s Represents the spatial attention map output by the fourth activation function layer; ||M s ||1 means M s The L1-norm of M s [i,j] represents M s The pixel in the i-th row and j-th column; H represents M c and M s Height; W means M c and M s Width.

[0128] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0129] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A remote sensing image aircraft positioning method based on a full-channel spatial attention module, characterized in that: The following steps are involved: S1: Acquire the remote sensing image to be located; S2: After preprocessing the remote sensing image to be located, it is input into a pre-trained remote sensing image aircraft positioning model to obtain the aircraft positioning result in the remote sensing image to be located; wherein the remote sensing image aircraft positioning model includes a convolutional neural network, a full-channel spatial attention module and a target detection and positioning module connected in sequence.

2. According to claim 1, a remote sensing image aircraft positioning method based on a full-channel spatial attention module is characterized in that: The convolutional neural network includes a convolution module / a plurality of sequentially connected convolution modules; wherein the convolution module includes a sequentially connected convolution layer, a nonlinear activation function layer, a pooling layer and a batch normalization layer.

3. The remote sensing image aircraft positioning method based on the full-channel spatial attention module according to claim 2 is characterized in that: The full-channel spatial attention module includes an adaptive average pooling layer, a first 1*1 convolution layer, a first activation function layer, a second 1*1 convolution layer, an adaptive maximum pooling layer, a third 1*1 convolution layer, a second activation function layer, a fourth 1*1 convolution layer, an element-by-element addition unit, a third activation function layer, a first element-by-element multiplication unit, a channel maximum value processing unit, a channel average value processing unit, a channel splicing layer, a 7*7 convolution layer, a fourth activation function layer, and a second element-by-element multiplication unit; The input ends of the adaptive average pooling layer and the adaptive maximum pooling layer are both connected to the output end of the convolutional neural network; The output end of the adaptive average pooling layer is connected to the input end of the element-by-element addition unit through the first 1*1 convolution layer, the first activation function layer, and the second 1*1 convolution layer in sequence; The output end of the adaptive maximum pooling layer is connected to the input end of the element-by-element addition unit through the third 1*1 convolution layer, the second activation function layer, and the fourth 1*1 convolution layer in sequence; The output end of the element-by-element addition unit is connected to the input end of the first element-by-element multiplication unit through the third activation function layer; The output end of the convolutional neural network is connected to the input end of the first element-by-element multiplication unit; The output end of the first element-by-element multiplication unit is connected to the input end of the channel maximum value processing unit and the channel average value processing unit respectively; The output ends of the channel maximum value processing unit and the channel average value processing unit are both connected to the input end of the channel splicing layer; The output end of the channel splicing layer is connected to the input end of the second element-by-element multiplication unit through the 7*7 convolution layer and the fourth activation function layer in sequence; An output terminal of the first element-by-element multiplication unit is connected to an input terminal of the second element-by-element multiplication unit; An output end of the second element-by-element multiplication unit is connected to an input end of the target detection and positioning module.

4. The remote sensing image aircraft positioning method based on the full-channel spatial attention module according to claim 3 is characterized in that: The target detection and positioning module includes a convolutional layer, a nonlinear activation layer, a downsampling layer and an upsampling layer connected in sequence; The input end of the convolutional layer of the target detection and positioning module is connected to the output end of the second element-by-element multiplication unit; The output end of the upsampling layer is the output end of the remote sensing image aircraft positioning model.

5. The remote sensing image aircraft positioning method based on the full-channel spatial attention module according to claim 1 is characterized in that: The preprocessing in S2 includes scaling, normalization and denoising.

6. The remote sensing image aircraft positioning method based on the full-channel spatial attention module according to claim 4 is characterized in that: The pre-trained remote sensing image aircraft localization model is obtained based on the following method: S21: Acquire an initial remote sensing image data set; wherein the initial remote sensing image data set includes a plurality of remote sensing images including aircraft; S22: annotating the initial remote sensing image data set; wherein the annotation content includes a bounding box of the aircraft, coordinates of the bounding box, and a category of the bounding box; S23: preprocessing the annotated initial remote sensing image data set to obtain a final remote sensing image data set; S24: Inputting the final remote sensing image data set into a remote sensing image aircraft positioning model to train it, thereby obtaining a pre-trained remote sensing image aircraft positioning model.

7. The remote sensing image aircraft positioning method based on the full-channel spatial attention module according to claim 6 is characterized in that: The preprocessing in S23 includes scaling processing, normalization processing, data enhancement processing and denoising processing.

8. The remote sensing image aircraft positioning method based on the full-channel spatial attention module according to claim 7 is characterized in that: The remote sensing image aircraft localization model is trained based on the following total loss function: L total =μL t +vL m +ηL s ; Among them, L total Represents the total loss; L t , L m , L s They represent target loss, feature enhancement consistency loss, and attention sparsity loss respectively; μ, v, η represent weight coefficients.

9. The remote sensing image aircraft positioning method based on the full-channel spatial attention module according to claim 8 is characterized by: L t =αL box +βL cls ; L s =γ1L channel +γ2L spatial ; Among them, L box represents the bounding box regression loss; L cls represents classification loss; α and β represent weight coefficients; X O Represents the original features of the convolutional neural network output; X O [i] indicates X O The i-th pixel of en represents the enhanced features output by the full-channel spatial attention module; X en [i] indicates X en The i-th pixel of O and X en The number of pixels included; L channel represents the channel attention sparsity loss; L spatial represents the spatial attention sparsity loss; γ1 and γ2 represent hyperparameters; ||X O [i]-X en [i]|| 2 Represents X O [i]-X en The square of the Euclidean norm of [i].

10. The remote sensing image aircraft positioning method based on the full-channel spatial attention module according to claim 9 is characterized in that: L box =1-GIoU(B true ,B pred ); L cls =-y c log(p c ); Among them, GIoU represents the overlap between the predicted bounding box and the true bounding box; B true represents the ground-truth bounding box of the annotation; B pred represents the predicted bounding box; y c One-hot encoding of the category representing the true bounding box; p c represents the classification probability of the predicted bounding box; M c Represents the channel attention map of the third activation function layer; ||M c ||1 means M c The L1-norm of M c [i,j] represents M c The pixel in the i-th row and j-th column; M s Represents the spatial attention map output by the fourth activation function layer; ||M s ||1 means M s The L1-norm of M s [i,j] represents M s The pixel in the i-th row and j-th column; H represents M c and M s Height; W means M c and M s Width.

Images