Remote sensing image segmentation method and system based on lightweight linear attention and multi-order feature aggregation

Abstract

The application discloses a kind of remote sensing image segmentation method and system based on lightweight linear attention and multi-order feature aggregation, comprising: S1.Acquire remote sensing image, construct lightweight linear attention feature extractor, capture the global information of remote sensing image by introducing linear layer and convolution layer, and the global and local features of remote sensing image are extracted by linear attention mechanism weighting feature;S2.Construct multi-order feature aggregation engine, the output feature of feature extractor is fused multi-scale feature information by multi-level step-by-step optimization, and extract high-dimensional semantic information in remote sensing image;S3.Based on the feature of multi-order feature aggregation engine output predicts the segmentation result of image, and constructs loss function optimization segmentation result, obtains final segmentation image;The application introduces lightweight linear attention mechanism and multi-order feature aggregation engine, efficiently extracts and fuses multi-scale feature in remote sensing image, while improving segmentation accuracy, reduces the consumption of computing resources.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing image processing, and more particularly to a remote sensing image segmentation method and system based on lightweight linear attention and multi-order feature aggregation. Background Art

[0002] Remote sensing image segmentation technology is one of the key tasks in remote sensing image analysis and is widely used in land use, cover classification, urban planning, environmental monitoring, and disaster assessment.

[0003] In recent years, with the continuous development of remote sensing technology, the resolution of acquired images has become increasingly higher, and the amount of data has also increased accordingly. However, remote sensing images often have complex backgrounds, objects of different scales, and high-dimensional spatial information. This makes remote sensing image segmentation face many challenges. In particular, when processing large-scale images and diverse scenes, traditional segmentation methods have difficulty meeting the requirements of high accuracy, real-time performance, and robustness.

[0004] At present, most of the existing remote sensing image segmentation methods rely on deep learning models, especially convolutional neural networks (CNNs) and self-attention mechanisms. Although these methods have improved the segmentation accuracy to a certain extent, there are still certain bottlenecks in computational complexity and model efficiency.

[0005] For example, the traditional self-attention mechanism captures global information by calculating the similarity between each pixel, but the time complexity of this calculation method increases quadratically with the increase of image size, resulting in low computational efficiency; in addition, in remote sensing images, targets usually have different scales and shapes, and traditional feature extraction methods often find it difficult to capture both global structure and local detail information at the same time. Especially when processing high-resolution remote sensing images, traditional methods often cannot achieve efficient calculation while ensuring accuracy.

[0006] Therefore, how to improve segmentation accuracy while reducing the consumption of computing resources in remote sensing image processing, especially in large-scale remote sensing image processing tasks, is an urgent problem that technicians in this field need to solve. Summary of the Invention

[0007] In view of this, the present invention provides a remote sensing image segmentation method and system based on lightweight linear attention and multi-order feature aggregation, aiming to improve segmentation accuracy while reducing the consumption of computing resources, especially showing stronger advantages in large-scale remote sensing image processing tasks.

[0008] In order to achieve the above object, the present invention adopts the following technical solutions:

[0009] A remote sensing image segmentation method based on lightweight linear attention and multi-order feature aggregation includes the following steps:

[0010] S1. Obtain a remote sensing image, construct a lightweight linear attention feature extractor, capture global information of the remote sensing image by introducing linear layers and convolutional layers, and weight the features by linear attention mechanisms to extract global and local features in the remote sensing image;

[0011] S2. Construct a multi-order feature aggregation engine to gradually optimize and fuse multi-scale feature information by multiple levels to extract high-dimensional semantic information in the remote sensing image from the output features of the lightweight linear attention feature extractor;

[0012] S3. Predict the segmentation result of the image based on the features output by the multi-order feature aggregation engine, and construct a loss function to optimize the segmentation result to obtain the final segmentation image.

[0013] Preferably, the specific content of step S1 includes:

[0014] S11. Perform preprocessing operations on the obtained remote sensing image to enhance image quality and remove noise, and input the preprocessed image into the feature extraction network to extract preliminary features through initial convolutional layers and normalization layers;

[0015] S12. The preliminary features are processed through linear layers, convolutional layers, and activation functions to generate corresponding queries Q, keys K, and values V, and linear attention mechanisms are used to calculate weighted features;

[0016] S13. The weighted features are processed through linear layers and activation functions to obtain enhanced features, which are then point-wise multiplied with the weighted features. After processing through linear layers, the enhanced features are added to the preliminary features to obtain attention-enhanced features;

[0017] S14. The attention-enhanced features are further optimized through normalization layers and multi-layer perceptrons, and are added to the attention-enhanced features to obtain the output features of the lightweight linear attention feature extractor.

[0018] Preferably, in step S12, the method for calculating weighted features using linear attention mechanisms is as follows:

[0019]

[0020] Where i represents the current sequence position or feature index, j is a loop variable used to traverse all elements or features in the sequence, and N represents the length of the sequence or the number of features.

[0021] Preferably, the specific content of step S2 is as follows:

[0022] ​S21. Process the output features of the lightweight linear attention feature extractor through normalization and convolution layers to obtain preliminary features. Then, perform dimensionality reduction through a global average pooling layer to extract global information of the image. Then, perform element-wise subtraction to calculate the deviation of each pixel relative to the global information to obtain differentiated features. These differentiated features are point-wise multiplied with the output features of the lightweight linear attention feature extractor to obtain first-stage features.

[0023] S22. The first stage features are added to the preliminary features of step S21 to obtain fused features. The fused features are processed by an activation function and further abstracted and optimized by a depthwise separable convolutional layer to obtain intermediate features.

[0024] S23. Cut the intermediate features into sub-feature maps of different dimensions. Process each sub-feature map through depthwise separable convolutions of different sizes to extract feature details at different scales. Concatenate and fuse the processed feature maps to obtain the final multi-scale features.

[0025] S24. After the multi-scale features and the first-stage features are processed by the convolution layer and the activation function respectively, they are multiplied point by point to obtain the output features of the multi-order feature aggregation engine.

[0026] Preferably, the specific contents of step S3 include:

[0027] S31. The output features of the multi-level feature aggregation engine are further processed through a convolutional layer and an activation function to obtain a preliminary segmentation probability map.

[0028] S32. performing post-processing operations on the preliminary segmentation probability map to optimize the segmentation result;

[0029] S33. Design a loss function to train the model and optimize the segmentation results to obtain the final segmented image.

[0030] Preferably, the loss function is a comprehensive loss function including cross entropy loss, Dice coefficient loss and IoU loss.

[0031] A remote sensing image segmentation system based on lightweight linear attention and multi-order feature aggregation, based on the remote sensing image segmentation method based on lightweight linear attention and multi-order feature aggregation, comprising: an image acquisition module, a lightweight linear attention feature extraction module, a multi-order feature aggregation engine module and an optimization output module;

[0032] Image acquisition module, used to acquire remote sensing images;

[0033] A lightweight linear attention feature extraction module is used to build a lightweight linear attention feature extractor. It captures the global information of remote sensing images by introducing linear layers and convolutional layers, and extracts global and local features from remote sensing images by weighting features through the linear attention mechanism.

[0034] The multi-level feature aggregation engine module is used to build a multi-level feature aggregation engine. It gradually optimizes the output features of the lightweight linear attention feature extractor and fuses multi-scale feature information through multiple levels to extract high-dimensional semantic information from remote sensing images.

[0035] The optimization output module is used to predict the image segmentation results based on the features output by the multi-order feature aggregation engine, and to construct a loss function to optimize the segmentation results to obtain the final segmented image.

[0036] Preferably, the lightweight linear attention feature extraction module includes an image preprocessing and preliminary feature extraction unit, a linear attention calculation unit, a feature weighting and enhancement unit, and a feature optimization and output unit;

[0037] The multi-order feature aggregation engine module includes a preliminary feature enhancement and weighting unit, a feature fusion and convolution enhancement unit, a multi-scale feature extraction and fusion unit, and a feature weighting and final output unit.

[0038] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements a remote sensing image segmentation method based on lightweight linear attention and multi-order feature aggregation.

[0039] A processing terminal includes a memory and a processor. The memory stores a computer program that can be run on the processor. When the processor executes the computer program, it implements the remote sensing image segmentation method based on lightweight linear attention and multi-order feature aggregation.

[0040] As can be seen from the above technical solutions, compared with the existing technology, the present invention discloses a remote sensing image segmentation method and system based on lightweight linear attention and multi-order feature aggregation. By introducing a lightweight linear attention mechanism and a multi-order feature aggregation engine, it can efficiently extract and fuse multi-scale features in remote sensing images, which not only improves the image segmentation accuracy but also optimizes the computational efficiency. It is suitable for the automated processing and efficient segmentation of large-scale remote sensing images.

[0041] Specifically: By introducing a lightweight linear attention mechanism, the computational complexity of the traditional self-attention mechanism is significantly reduced, and the computational efficiency is improved, making it particularly suitable for the segmentation of large-scale remote sensing images. At the same time, the use of efficient operations such as depthwise separable convolution and global average pooling gives this method strong parallel computing capabilities, making full use of modern hardware acceleration to improve processing speed.

[0042] The multi-order feature aggregation engine enhances the perception of complex backgrounds and details by gradually optimizing and fusing features at different scales, showing significant accuracy improvements in the segmentation of targets such as ships and roads in remote sensing images. At the same time, the strategy of combining global information with local details improves the adaptability of the model in different scenarios and ensures the robustness of segmentation accuracy in diverse and complex environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0044] Figure 1 A schematic diagram of a remote sensing image segmentation method based on lightweight linear attention and multi-order feature aggregation provided by the present invention;

[0045] Figure 2 Schematic diagram of the lightweight linear attention feature extractor provided by the present invention;

[0046] Figure 3 Schematic diagram of the multi-level feature aggregation engine provided by the present invention. DETAILED DESCRIPTION

[0047] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. 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 making creative efforts are within the scope of protection of the present invention.

[0048] The embodiment of the present invention discloses a remote sensing image segmentation method based on lightweight linear attention and multi-order feature aggregation, such as Figure 1 , including the following steps:

[0049] S1. Obtain remote sensing images and construct a lightweight linear attention feature extractor. This extractor captures the global information of the remote sensing image by introducing linear and convolutional layers. It also weights the features using a linear attention mechanism to extract global and local features from the remote sensing image.

[0050] S2. Build a multi-level feature aggregation engine to extract high-dimensional semantic information from remote sensing images by optimizing the output features of the lightweight linear attention feature extractor and fusing multi-scale feature information through multiple levels.

[0051] S3. Predict the image segmentation results based on the features output by the multi-level feature aggregation engine, and construct a loss function to optimize the segmentation results to obtain the final segmented image.

[0052] In order to further implement the above technical solutions, Figure 2 , the specific contents of step S1 include:

[0053] S11. Image preprocessing and preliminary feature extraction: The acquired remote sensing images are preprocessed to enhance image quality and remove noise. The preprocessed images are input into the feature extraction network, and preliminary features are extracted through the initial convolution layer and normalization layer.

[0054] In this embodiment, a series of pre-processing operations are performed on the image, including but not limited to image size normalization, color space conversion, etc., to enhance image quality and remove noise;

[0055] S12. Linear Attention Computation: Preliminary Features Through different operations of linear layer, convolution layer and activation function, the corresponding query Q, key K and value V are generated, and the weighted features are calculated using the linear attention mechanism.

[0056] S13. Feature Weighting and Enhancement: Weighted Features After the linear layer and activation function, the enhanced features are obtained And with weighted features Perform point-by-point multiplication and process it through the linear layer to combine it with the initial features Add together to get the attention enhancement feature

[0057]

[0058] in, Represents matrix addition, ⊙ represents point-by-point multiplication, and Linear represents a linear layer;

[0059] feature Able to effectively fuse global information from different scales;

[0060] S14. Feature Optimization and Output: Enhance Attention Features The optimized features are further optimized through the normalization layer BN and multi-layer perceptron MLP, and the optimized features are combined with the attention enhancement features. Add and get the output features of the lightweight linear attention feature extractor

[0061]

[0062] In this embodiment, the normalization layer helps to accelerate the training process and improve stability, while the multilayer perceptron further enhances the expressiveness of features through nonlinear transformation;

[0063] Through multi-level feature fusion and optimization, the final feature It provides rich semantic information for subsequent image segmentation tasks, thereby improving the accuracy and robustness of remote sensing image segmentation.

[0064] In order to further implement the above technical solution, in step S12, the linear attention mechanism is used to calculate the weighted features. The method is:

[0065]

[0066] Among them, i represents the current sequence position or feature index. When processing sequence data, i usually represents the i-th element or feature; j is a loop variable used to traverse all elements or features in the sequence. In the summation operation, j changes from 1 to N, where N represents the length of the sequence or the number of features. It is the total number of elements or features in the sequence. The linear attention mechanism is used to calculate the weighted features. The method changes the order of attention calculation from the traditional (QK T )V to Q(K T V), effectively reducing the computational complexity from O(N 2 ) is reduced to O(N), greatly improving the computational efficiency.

[0067] In practical applications, the advantages of linear attention include:

[0068] Efficiency: It reduces computational complexity and is suitable for processing long sequence data in remote sensing images, which is especially important in large-scale image segmentation tasks.

[0069] Global modeling capabilities: Each query can aggregate key and value information from the entire image, helping to capture global context;

[0070] Scalability: Suitable for large-scale parallel computing and able to take advantage of modern hardware acceleration to achieve efficient image processing.

[0071] The lightweight linear attention feature extractor effectively combines the global and local information of remote sensing images and is particularly suitable for the semantic segmentation of targets such as ships and roads in complex backgrounds. Through the lightweight linear attention mechanism, it not only ensures the efficiency of calculation but also improves the accuracy of feature extraction, making this method have strong application potential in remote sensing image segmentation tasks.

[0072] In order to further implement the above technical solutions, Figure 3 , the specific content of step S2 is:

[0073] S21. Output features of lightweight linear attention feature extractor After the normalization layer and the convolution layer processing, the preliminary feature φ is obtained, and the dimension is reduced by the global average pooling layer GAP to extract the global information of the image and calculate the deviation of each pixel relative to the global information through element-by-element subtraction operation to obtain the differential feature. The differential feature is combined with the output feature of the lightweight linear attention feature extractor Perform point-by-point multiplication to obtain the first-stage feature φ1;

[0074]

[0075] Among them, GAP represents the global average pooling layer, and ⊙ represents the point-by-point multiplication operation, which can enhance the detail information, highlight the important areas in the image, and effectively improve the expressiveness of the features;

[0076] S22. The first-stage feature φ1 is added to the preliminary feature φ in step S21 to obtain a fused feature. The fused feature is processed by an activation function and further abstracted and optimized by a depthwise separable convolutional layer to obtain an intermediate feature φ2.

[0077] In this embodiment, the depthwise separable convolution layer is a 5×5 depthwise separable convolution layer. The depthwise separable convolution can effectively reduce the computational complexity while maintaining a good feature extraction effect.

[0078] S23. Cut the intermediate feature φ2 into sub-feature maps of different dimensions. Process each sub-feature map through depthwise separable convolutions of different sizes to extract feature details at different scales. Concatenate and fuse the processed feature maps to obtain the final multi-scale feature φ3. This feature map integrates information from different scales and can more comprehensively represent the diverse features in the image.

[0079] The intermediate feature φ2 is specifically divided into and Represent feature information of different scales respectively, and each sub-feature map is processed by depth-wise separable convolution of different sizes: the dimension is The sub-features of are processed by a 5×5 depth-wise separable convolutional layer with a dimension of The sub-features of are processed by a 7×7 depthwise separable convolutional layer. This operation helps to extract feature details at different scales and enhance the perception of different object scales.

[0080] S24. The multi-scale feature φ3 is processed by a convolution layer and an activation function SiLU respectively, and then multiplied point by point to obtain the output feature of the multi-order feature aggregation engine

[0081] In this embodiment, the output feature of the multi-order feature aggregation engine is

[0082]

[0083] where SiLU represents the SiLU activation function, and Conv represents the convolution layer. In this way, the final feature aggregates multi-scale information and enhances the fine-grained segmentation capability in image segmentation tasks, especially for complex backgrounds and details in remote sensing images.

[0084] The multi-order feature aggregation engine can effectively improve the accuracy and robustness of remote sensing image segmentation by step-by-step optimization and fusion of multi-scale feature information, and is particularly suitable for target extraction and segmentation in complex scenes. By using techniques such as depth separable convolution and global average pooling, not only the computational efficiency is improved, but also the model's perception of various details and global structures in images is enhanced.

[0085] To further implement the above technical solutions, the specific content of step S3 includes:

[0086] S31. The output feature of the multi-order feature aggregation engine is further processed by a convolution layer and an activation function to ensure that the feature has an appropriate expression form for segmentation, obtaining a preliminary segmentation probability map;

[0087] The output of the convolution layer is the predicted value of each pixel, representing the probability value of the pixel belonging to each class. By applying a suitable activation function (such as Softmax or Sigmoid), the predicted value is converted into a probability distribution, which represents the class probability of each pixel, This process converts the feature map into an actual segmentation probability map as the output of image segmentation.

[0088] S32. The preliminary segmentation probability map is post-processed to optimize the segmentation result;

[0089] In practical applications, thresholding operations can be used to convert the probability of each pixel into a class label, or conditional random fields (CRF) can be used for refinement to remove noise in the prediction and enhance edges. The specific post-processing operation depends on the task requirements and can be determined through experiments to determine the most appropriate method.

[0090] ​S33. Design a loss function to train the model and optimize the segmentation result, and obtain a final segmentation image.

[0091] To further implement the above technical solution, the loss function is a comprehensive loss function including cross-entropy loss, Dice coefficient loss and IoU loss.

[0092]

[0093] Wherein, a, β and γ are weight coefficients, used to adjust the influence of each loss term; in the training process, the performance of the model on different loss terms can be balanced by adjusting the weight coefficients, and finally the best segmentation effect is achieved; the final segmentation image can not only accurately segment the target, but also has strong adaptability when facing complex background and different scale targets;

[0094] In this embodiment, the cross-entropy loss is suitable for pixel-level classification task, which measures the difference between the predicted class probability and the true label of the model, and is specifically defined as:

[0095]

[0096] Wherein, y a，c is the true label (one-hot encoding) of the a-th pixel, is the predicted probability of the a-th pixel in class c;

[0097] The Dice coefficient loss is used to measure the overlap degree of the segmentation result, and is particularly suitable for handling class imbalance problem, and the Dice coefficient loss is defined as:

[0098]

[0099] Wherein, and y i are the prediction and true label of the a-th pixel respectively, and the closer the Dice coefficient is to 1, the better the segmentation result is;

[0100] The IoU loss is used to measure the overlap degree between the predicted segmentation result and the true label, and the definition of the IoU loss is:

[0101]

[0102] Wherein, IoU is the intersection over union of the predicted and true segmentation regions, and the higher the IoU value is, the better the segmentation effect is.

[0103] A remote sensing image segmentation system based on lightweight linear attention and multi-order feature aggregation is based on a remote sensing image segmentation method based on lightweight linear attention and multi-order feature aggregation, comprising: an image acquisition module, a lightweight linear attention feature extraction module, a multi-order feature aggregation engine module and an optimization output module;

[0104] Image acquisition module, used to acquire remote sensing images;

[0105] A lightweight linear attention feature extraction module is used to build a lightweight linear attention feature extractor. It captures the global information of remote sensing images by introducing linear layers and convolutional layers, and extracts global and local features from remote sensing images by weighting features through the linear attention mechanism.

[0106] The multi-level feature aggregation engine module is used to build a multi-level feature aggregation engine. It gradually optimizes the output features of the lightweight linear attention feature extractor and fuses multi-scale feature information through multiple levels to extract high-dimensional semantic information from remote sensing images.

[0107] The optimization output module is used to predict the image segmentation results based on the features output by the multi-order feature aggregation engine, and to construct a loss function to optimize the segmentation results to obtain the final segmented image.

[0108] To further implement the above technical solution, the lightweight linear attention feature extraction module includes an image preprocessing and preliminary feature extraction unit, a linear attention calculation unit, a feature weighting and enhancement unit, and a feature optimization and output unit;

[0109] The multi-order feature aggregation engine module includes a preliminary feature enhancement and weighting unit, a feature fusion and convolution enhancement unit, a multi-scale feature extraction and fusion unit, and a feature weighting and final output unit.

[0110] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements a remote sensing image segmentation method based on lightweight linear attention and multi-order feature aggregation.

[0111] A processing terminal includes a memory and a processor. The memory stores a computer program that can be run on the processor. When the processor executes the computer program, a remote sensing image segmentation method based on lightweight linear attention and multi-order feature aggregation is implemented.

[0112] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

[0113] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily 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 is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A remote sensing image segmentation method based on lightweight linear attention and multi-order feature aggregation, characterized by: The following steps are involved: S1. Obtain remote sensing images and construct a lightweight linear attention feature extractor. This extractor captures the global information of the remote sensing image by introducing linear and convolutional layers. It also weights the features using a linear attention mechanism to extract global and local features from the remote sensing image. S2. Build a multi-level feature aggregation engine to extract high-dimensional semantic information from remote sensing images by optimizing the output features of the lightweight linear attention feature extractor and fusing multi-scale feature information through multiple levels. S3. Predict the image segmentation result based on the features output by the multi-level feature aggregation engine, and construct a loss function to optimize the segmentation result to obtain the final segmented image; The specific contents of step S1 include: S11. Preprocess the acquired remote sensing image to enhance image quality and remove noise. The preprocessed image is input into the feature extraction network, which extracts preliminary features through the initial convolutional layer and normalization layer. S12. The preliminary features are processed through different operations such as linear layers, convolutional layers, and activation functions to generate the corresponding query Q, key K, and value V. The weighted features are calculated using the linear attention mechanism. S13. The weighted features are processed through a linear layer and an activation function to obtain enhanced features, which are then point-wise multiplied with the weighted features. After processing through a linear layer, they are added to the preliminary features to obtain attention-enhanced features. S14. The attention-enhanced feature is further optimized through a normalization layer and a multi-layer perceptron, and is added to the attention-enhanced feature to obtain the output feature of the lightweight linear attention feature extractor; The specific content of step S2 is: S21. Process the output features of the lightweight linear attention feature extractor through normalization and convolution layers to obtain preliminary features. Then, perform dimensionality reduction through a global average pooling layer to extract global information of the image. Then, perform element-wise subtraction to calculate the deviation of each pixel relative to the global information to obtain differentiated features. These differentiated features are point-wise multiplied with the output features of the lightweight linear attention feature extractor to obtain first-stage features. S22. The first stage features are added to the preliminary features of step S21 to obtain fused features. The fused features are processed by an activation function and further abstracted and optimized by a depthwise separable convolutional layer to obtain intermediate features. S23. Cut the intermediate features into sub-feature maps of different dimensions. Process each sub-feature map through depthwise separable convolutions of different sizes to extract feature details at different scales. Concatenate and fuse the processed feature maps to obtain the final multi-scale features. S24. After the multi-scale features and the first-stage features are processed by the convolution layer and the activation function respectively, they are multiplied point by point to obtain the output features of the multi-order feature aggregation engine.

2. The remote sensing image segmentation method based on lightweight linear attention and multi-order feature aggregation according to claim 1, characterized in that: In step S12, the weighted features are calculated using the linear attention mechanism The method is: ； Where i represents the current sequence position or feature index, j is a loop variable used to traverse all elements or features in the sequence, and N represents the length of the sequence or the number of features.

3. The remote sensing image segmentation method based on lightweight linear attention and multi-order feature aggregation according to claim 1, characterized in that: The specific contents of step S3 include: S31. The output features of the multi-level feature aggregation engine are further processed through a convolutional layer and an activation function to obtain a preliminary segmentation probability map. S32. performing post-processing operations on the preliminary segmentation probability map to optimize the segmentation result; S33. Design a loss function to train the model and optimize the segmentation results to obtain the final segmented image.

4. The remote sensing image segmentation method based on lightweight linear attention and multi-order feature aggregation according to claim 1, characterized in that: The loss function is a comprehensive loss function including cross entropy loss, Dice coefficient loss and IoU loss.

5. A remote sensing image segmentation system based on lightweight linear attention and multi-order feature aggregation, characterized by: A remote sensing image segmentation method based on lightweight linear attention and multi-order feature aggregation according to any one of claims 1 to 4, comprising: an image acquisition module, a lightweight linear attention feature extraction module, a multi-order feature aggregation engine module, and an optimization output module; Image acquisition module, used to acquire remote sensing images; A lightweight linear attention feature extraction module is used to build a lightweight linear attention feature extractor. It captures the global information of remote sensing images by introducing linear layers and convolutional layers, and extracts global and local features from remote sensing images by weighting features through the linear attention mechanism. The multi-level feature aggregation engine module is used to build a multi-level feature aggregation engine. It gradually optimizes the output features of the lightweight linear attention feature extractor and fuses multi-scale feature information through multiple levels to extract high-dimensional semantic information from remote sensing images. The optimization output module is used to predict the image segmentation results based on the features output by the multi-order feature aggregation engine, and to construct a loss function to optimize the segmentation results to obtain the final segmented image.

6. The remote sensing image segmentation system based on lightweight linear attention and multi-order feature aggregation according to claim 5, characterized in that: The lightweight linear attention feature extraction module includes an image preprocessing and preliminary feature extraction unit, a linear attention calculation unit, a feature weighting and enhancement unit, and a feature optimization and output unit; The multi-order feature aggregation engine module includes a preliminary feature enhancement and weighting unit, a feature fusion and convolution enhancement unit, a multi-scale feature extraction and fusion unit, and a feature weighting and final output unit.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, it implements the remote sensing image segmentation method based on lightweight linear attention and multi-order feature aggregation as described in any one of claims 1 to 4.

8. A processing terminal comprising a memory and a processor, wherein the memory stores a computer program that can be run on the processor, characterized in that: When the processor executes the computer program, it implements a remote sensing image segmentation method based on lightweight linear attention and multi-order feature aggregation as described in any one of claims 1 to 4.

Images