A remote sensing image segmentation method combined with similarity attention without additional parameters

Abstract

The present invention discloses a remote sensing image segmentation method that combines parameter-free similarity attention. The method relates to the field of remote sensing image segmentation and includes the following steps: utilizing a feature extraction and fusion optimization network to extract multi-scale features of the remote sensing image and perform fusion optimization to generate initial features; utilizing a parameter-free similarity attention unit in a feature enhancement network to enhance sensitivity, and utilizing a boundary and texture feature extraction unit to extract boundary and texture features, which are then fused to generate enhanced features; and utilizing a segmentation network to optimize and predict the categories of the enhanced features, outputting the segmentation results of the remote sensing image. This method significantly improves model efficiency without adding additional parameters, effectively addresses the issues of detail loss and boundary breakage in complex scenes, and improves adaptability and segmentation quality for low-contrast areas and high-noise remote sensing images.

Description

Technical Field

[0001] The present invention relates to the field of remote sensing image segmentation, and more particularly to a remote sensing image segmentation method combining similarity attention without additional parameters. Background Art

[0002] Remote sensing image segmentation is widely used in remote sensing image processing, including object classification, change detection, environmental monitoring, and urban planning. With the rapid development of remote sensing technology, image resolution continues to increase, and scene complexity has significantly increased, placing higher demands on the accuracy, detail recovery capabilities, and computational efficiency of segmentation algorithms.

[0003] However, traditional segmentation methods have significant limitations when dealing with complex remote sensing scenes. Highly complex backgrounds, small objects, and fuzzy boundary features make it difficult for traditional algorithms to accurately distinguish between object types. This is especially true in areas with detailed information, such as farmland edges and building outlines, where segmentation breaks or misclassifications are common. While deep learning-based convolutional neural networks improve segmentation performance through multi-layer feature extraction, their ability to capture global context remains insufficient, making it difficult to balance the representation of local details with the overall structure.

[0004] While existing attention-based methods can improve segmentation through adaptive feature enhancement, they generally rely on additional trainable parameters to adjust attention weights. This not only increases the model's parameter count and computational complexity, but also easily leads to overfitting due to parameter redundancy, limiting its application in remote sensing tasks with high real-time requirements. Furthermore, existing attention mechanisms lack robustness in low-contrast areas and under noise interference, resulting in the loss of critical boundary information.

[0005] On the other hand, existing methods still face challenges in detail recovery and noise suppression. High noise, uneven illumination, and blurred textures, common in high-resolution remote sensing images, make it difficult for traditional segmentation networks to effectively recover detailed information, resulting in blurred or broken edges in segmentation results. Although diffusion models have demonstrated strong detail recovery capabilities in image generation, their integration into remote sensing segmentation tasks is still immature, and existing methods lack systematic optimization strategies to balance detail accuracy and global consistency.

[0006] Therefore, how to design a remote sensing image segmentation method that combines similarity attention without additional parameters, which can effectively solve the problems of detail loss and boundary breakage in complex scenes, and improve the adaptability and segmentation quality of low-contrast areas and high-noise remote sensing images 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 combined with parameter-free similarity attention, which dynamically enhances boundary and texture features by constructing an adaptive attention unit that does not require additional parameters; combined with dilated convolution, depthwise separable convolution and diffusion model, it realizes efficient extraction of multi-scale features and detail restoration, significantly improving the segmentation accuracy and robustness in complex scenes while reducing computational complexity.

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

[0009] A remote sensing image segmentation method combined with similarity attention without additional parameters, comprising the following steps:

[0010] S1. Using feature extraction and fusion optimization network, extract multi-scale features of remote sensing images and perform fusion optimization to generate initial features;

[0011] S2. Based on the initial features, sensitivity enhancement is performed by the parameter-free similarity attention unit in the feature enhancement network, and boundary and texture feature extraction unit is used to extract boundary and texture features, and the enhanced features are generated by fusing them;

[0012] S3. Combined with the segmentation network, the enhanced features are optimized and the category prediction is performed, and the segmentation result of the remote sensing image is output.

[0013] Furthermore, the S1 includes:

[0014] S11. Use the Mamba pre-trained model to extract features from remote sensing images and obtain high-level semantic features φ;

[0015] S12. Based on the high-level semantic feature φ, multi-scale feature extraction is performed through the multi-scale convolution layer to obtain a multi-level feature φ containing global information. multi ;

[0016] S13, for multi-level features φ multi Perform weighted summation or concatenation fusion to generate comprehensive features φ fused ;

[0017] S14, through the convolution layer and activation function to the comprehensive feature φ fused Perform dimensionality reduction optimization to obtain the initial feature φ 0 .

[0018] Furthermore, in S12, the multi-scale convolution layer includes at least three convolution kernels of different sizes, and the multi-level feature φ multi Calculated by the following formula:

[0019] φ multi =[Conv k1 , Convk2 , Conv k2 ]

[0020] Among them, Conv k1 、Conv k2 、Conv k2 They represent convolution operations with kernel sizes of k1, k2, and k3, respectively, and [·] represents a concatenation operation.

[0021] Furthermore, in S2, the sensitivity enhancement is performed by the similarity attention unit without additional parameters, including:

[0022] S21, for the initial feature φ 0 Perform normalization to obtain normalized features

[0023] S22, based on normalized features Calculate the mean and variance of each neuron to assess its relative difference from other neurons; where a neuron represents the eigenvalue corresponding to any position or channel;

[0024] S23, through the energy function Quantify the neuronal differences and generate the energy matrix E;

[0025] S24, calculate the neuron importance weight according to the energy matrix E, and dynamically adjust the feature response through the Sigmoid function;

[0026] S25. Multiply the adjusted features with the original features to generate attention-enhanced features

[0027]

[0028] Where sigmoid(·) represents the sigmoid activation function.

[0029] Furthermore, in S23, the energy function Expressed as:

[0030]

[0031] Among them, t represents the current neuron, represents the mean of the remaining neurons after removing the current neuron, represents the variance of the remaining neurons after removing the current neuron, and λ represents the regularization parameter.

[0032] Furthermore, in S2, the boundary and texture feature extraction unit performs boundary and texture feature extraction, including:

[0033] The initial feature φ 0Input the hole convolution layer to expand the receptive field, and generate the boundary feature φ through the LeakyReLU activation function, convolution layer and ReLU activation function D ;

[0034] φ D =ReLU(Conv(LeakyReLU(DConv(φ 0 ))))

[0035] Among them, Conv(·) represents the convolutional layer operation, and DConv(·) represents the dilated convolutional layer operation;

[0036] And, the initial feature φ 0 Input the depth separable convolution layer and process it through the Mish activation function and the fully connected layer to generate the texture feature φ DW ;

[0037] φ DW =FC(Mish(DWConv(φ 0 )))

[0038] Among them, FC(·) represents the fully connected layer operation, and DWConv(·) represents the depthwise separable convolutional layer operation.

[0039] Furthermore, it also includes:

[0040] The boundary feature φ D and texture features φ D W splicing, obtain comprehensive features φ DDW ;

[0041] For comprehensive feature φ DDW Convolution, batch normalization and ReLU activation are performed in sequence, and the boundary feature φ D and texture features φ DW Fusion to obtain enhanced features

[0042] Furthermore, the S2 further includes:

[0043] Enhance the features Processed by ReLU activation function and enhanced with attention features after convolution and batch normalization Perform fusion to obtain fusion features;

[0044] The fusion feature is pixel-rearranged to obtain the enhanced feature φ + ;

[0045]

[0046] Among them, PShuffle(·) represents the pixel shuffle operation, and BN(·) represents the batch normalization operation.

[0047] Furthermore, the S3 includes:

[0048] S31, for the enhanced feature φ + Initial optimization is performed through convolutional layers, batch normalization layers, and activation functions to improve feature expression capabilities;

[0049] S32, based on the enhanced features after initial optimization, performing multi-step denoising processing through a diffusion model to gradually restore detail information and obtain optimized enhanced features;

[0050] S33. Combined with the segmentation head, the optimized enhanced features are converted into pixel-level classification results.

[0051] Furthermore, the segmentation network adopts a composite loss function L total ;

[0052] L total =λ1L CE +λ2L Dice +λ3L Boundary

[0053] Among them, λ1, λ2, λ3 are weight coefficients, L CE represents the cross entropy loss function, L Dice represents the Dice loss function, L Boundary represents the boundary loss function.

[0054] It can be seen from the above technical solution that compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0055] 1. This solution constructs a parameter-free similarity attention unit and utilizes normalization, mean squared error calculation, and an energy function to dynamically evaluate feature differences, eliminating the need for additional trainable parameters. Compared to traditional attention mechanisms, this significantly reduces the model's parameter count and computational complexity while maintaining the ability to adaptively focus on key image regions. In complex scenes of remote sensing imagery, it effectively enhances boundaries and details while avoiding the risk of overfitting due to parameter redundancy, thereby improving the real-time and robustness of segmentation tasks.

[0056] 2. Dilated convolution further expands the receptive field to capture global boundary information, combined with depthwise separable convolution to extract local texture details. The two types of features are then fused to form a comprehensive feature representation. Dilated convolution maintains resolution while covering a wider context range, while depthwise separable convolution reduces computational effort through channel separation. The two complement each other to optimize feature extraction capabilities for complex backgrounds and small objects. Furthermore, by fusing attention-enhanced features through pixel rearrangement, the model's adaptability to low-contrast areas and high-noise scenes in remote sensing images is further improved, ensuring the detail integrity and boundary accuracy of the segmentation results.

[0057] 3. A diffusion model is introduced at the back end of the segmentation network. This multi-step reverse denoising process gradually refines the feature map, restoring details lost due to noise or blur. This process combines a multi-objective optimization approach using cross-entropy, Dice, and boundary loss to not only enhance foreground and background classification accuracy but also improve the sharpness of segmentation edges through boundary constraints. Especially when processing high-resolution remote sensing imagery, the diffusion model effectively addresses the issues of blurred details and broken boundaries encountered in traditional methods, significantly improving segmentation clarity and overall consistency in complex scenes. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] 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.

[0059] Figure 1 A flow chart of a remote sensing image segmentation method combined with similarity attention without additional parameters provided by an embodiment of the present invention;

[0060] Figure 2 A schematic diagram of a process for generating enhanced features based on a feature enhancement network according to an embodiment of the present invention;

[0061] Figure 3 A schematic diagram of a process for outputting segmentation results based on a segmentation network according to an embodiment of the present invention;

[0062] Figure 4 A schematic diagram of the urban remote sensing image segmentation process combined with similarity attention without additional parameters provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0063] 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.

[0064] Example 1;

[0065] like Figure 1 As shown, this embodiment provides a remote sensing image segmentation method combined with similarity attention without additional parameters, including the following steps:

[0066] S1. Using feature extraction and fusion optimization network, extract multi-scale features of remote sensing images and perform fusion optimization to generate initial features;

[0067] S2. Based on the initial features, sensitivity enhancement is performed by the parameter-free similarity attention unit in the feature enhancement network, and boundary and texture feature extraction unit is used to extract boundary and texture features, and the enhanced features are generated by fusing them;

[0068] S3. Combined with the segmentation network, the enhanced features are optimized and the category prediction is performed, and the segmentation result of the remote sensing image is output.

[0069] This method significantly improves model efficiency through a highly efficient attention mechanism without additional parameters. It also integrates multiple techniques, such as dilated convolution and depthwise separable convolution, to enhance feature expression. It also introduces a diffusion model into the segmentation network to restore and optimize detail. This approach not only improves the accuracy and robustness of remote sensing image segmentation, but also significantly reduces computational complexity, effectively addressing the issues of blurred boundaries and lost detail in complex scenes.

[0070] The following further describes each step in the above method in detail;

[0071] In this embodiment S1, a feature extraction and fusion optimization network is used to extract multi-scale features of remote sensing images and perform fusion optimization to generate initial features. Specifically, the following steps are included:

[0072] S11. Use the Mamba pre-trained model to extract features from remote sensing images and obtain high-level semantic features φ. The Mamba model is trained on a large-scale remote sensing dataset. Its deep network structure excels at capturing complex patterns in remote sensing images. It can efficiently extract high-quality features from remote sensing images, providing a solid foundation for subsequent feature enhancement and segmentation tasks.

[0073] S12. Based on the high-level semantic feature φ, multi-scale feature extraction is performed through the multi-scale convolution layer to obtain a multi-level feature φ containing global information.multi ;

[0074] The multi-scale convolution layer includes at least three convolution kernels of different sizes, and the multi-level feature φ multi Calculated by the following formula:

[0075] φ multi =[Conv k1 , Conv k2 , Conv k2 ]

[0076] Among them, Conv k1 、Conv k2 、Conv k2 They represent convolution operations with kernel sizes of k1, k2, and k3, respectively, and [·] represents a concatenation operation.

[0077] Specifically, it can use 3×3, 5×5, and 7×7 convolution kernels to extract local, mid-range, and global features respectively, and integrate them into multi-level features through splicing operations; the multi-scale design covers ground objects of different scales in remote sensing images, avoids feature omissions caused by a single-scale convolution kernel, and enhances the model's generalization ability for complex scenes;

[0078] S13, for multi-level features φ multi Perform weighted summation or concatenation fusion to generate comprehensive features φ fused ; It can give adaptive weights to features of different scales to avoid redundancy;

[0079] S14, through the convolution layer and activation function to the comprehensive feature φ fused Perform dimensionality reduction optimization to reduce the complexity of subsequent calculations and obtain the initial feature φ 0 .

[0080] In this step, the use of the Mamba pre-trained model significantly improves the efficiency and accuracy of feature extraction. The design of the multi-scale convolutional layer enables the model to capture multi-scale information in the image, enhancing its adaptability to complex scenarios. Further, the comprehensive features generated through weighted summation or concatenation fusion retain global information while incorporating rich details, providing comprehensive feature support for subsequent feature enhancement and segmentation tasks.

[0081] like Figure 2 As shown, in this embodiment S2, based on the initial features, sensitivity enhancement is performed by the similarity attention unit without additional parameters in the feature enhancement network, and boundary and texture feature extraction unit performs boundary and texture feature extraction, and the enhanced features are generated by fusion;

[0082] Traditional attention mechanisms rely on trainable fully connected layers or convolutional layers to generate weights. In this embodiment, an energy function is used to directly quantify feature differences, avoiding parameter redundancy and reducing the risk of overfitting. No additional parameter similarity attention unit is used to enhance sensitivity, specifically including:

[0083] S21, for the initial feature φ 0 Perform normalization to obtain normalized features This operation ensures that the mean and variance of subsequent calculations are not affected by the feature dimension, improving the stability of the attention weight;

[0084] S22, based on normalized features Calculate the mean and variance of each neuron to assess its relative difference from other neurons; where a neuron represents the eigenvalue corresponding to any position or channel;

[0085] Specifically, for neuron t, the mean and variance of the remaining neurons after removing neuron t are calculated. By excluding the current neuron t, the mean and variance of the remaining features are calculated to reflect the relative relationship between this neuron and the overall features. If the feature value of a neuron is much higher than that of other neurons, then after removing this value, the mean and variance of the remaining features will drop significantly, thus reflecting its uniqueness. By excluding itself, the relative difference between the current neuron and the overall features can be accurately assessed, avoiding interference from autocorrelation.

[0086] S23, through the energy function Quantify the neuron differences and generate the energy matrix E; energy function Expressed as:

[0087]

[0088] Among them, t represents the current neuron, represents the mean of the remaining neurons after removing the current neuron, represents the variance of the remaining neurons after removing the current neuron, and λ represents the regularization parameter;

[0089] In the specific energy function In , the coefficient 4 is used to adjust the overall range of energy values ​​so that they are distributed in a more reasonable range; The purpose is to normalize the energy value to ensure that the function is insensitive to scale. It is used to measure the degree of deviation between the current neuron and the mean of the remaining neurons. The greater the deviation, the higher the energy value. The purpose is to balance local differences and overall stability by introducing the variance and regularization terms of the remaining features. This function objectively evaluates the importance of features through statistical methods, avoiding subjective parameter setting. The regularization parameter λ enhances numerical stability and is suitable for high-noise remote sensing scenarios.

[0090] S24, calculate the neuron importance weight according to the energy matrix E, and dynamically adjust the feature response through the Sigmoid function;

[0091] S25. Multiply the adjusted features with the original features to generate attention-enhanced features

[0092]

[0093] Here, sigmoid(·) represents the sigmoid activation function. The sigmoid activation function is used to smoothly adjust the importance of each neuron, so that the activation value of each neuron is dynamically adjusted according to its relative importance, thereby improving the attention to details and boundary information.

[0094] Furthermore, the boundary and texture feature extraction unit performs boundary and texture feature extraction, including:

[0095] The initial feature φ 0 Input the hole convolution layer to expand the receptive field, and generate the boundary feature φ through the LeakyReLU activation function, convolution layer and ReLU activation function D Specifically, it uses a 3×3 convolution kernel with a dilation rate of r=2 to expand the receptive field to (2r+1)×(2r+1), capturing long-distance boundary information without reducing the resolution.

[0096] φ D =ReLU(Conv(LeakyReLU(DConv(φ 0 ))))

[0097] Among them, Conv(·) represents the convolution layer operation, and DConv(·) represents the dilated convolution layer operation;

[0098] And, the initial feature φ 0 Input the depth separable convolution layer and process it through the Mish activation function and the fully connected layer to generate the texture feature φ DW Specifically, the standard convolution is decomposed into a channel-by-channel convolution with a 3×3 kernel and a point-by-point convolution with a 1×1 kernel, reducing the number of parameters to 1 / N, thus reducing the amount of computation while preserving texture details.

[0099] φ DW =FC(Missh(DWConv(φ 0 )))

[0100] Among them, FC(·) represents the fully connected layer operation, and DWConv(·) represents the depthwise separable convolutional layer operation.

[0101] The dilated convolution here focuses on the global boundary, and the depthwise separable convolution extracts the local texture. The two achieve feature complementarity through splicing and optimization, solving the problem that a single convolution kernel is difficult to take into account both long- and short-range dependencies.

[0102] Furthermore, it also includes:

[0103] The boundary feature φ D and texture features φ DW Splicing to obtain comprehensive features φ DDW ;

[0104] For comprehensive feature φ DDW Convolution, batch normalization and ReLU activation are performed in sequence, and the boundary feature φ D and texture features φ DW Fusion to obtain enhanced features The splicing operation integrates boundary and texture information, and the convolutional layer further optimizes the feature representation, enhancing the model's ability to segment complex objects.

[0105] Furthermore, the above steps also include:

[0106] Enhance the features Processed by ReLU activation function and enhanced with attention features after convolution and batch normalization Perform fusion to obtain fusion features;

[0107] The fusion feature is pixel-rearranged to obtain the enhanced feature φ + ;

[0108]

[0109] Among them, PShuffle(·) represents the pixel rearrangement operation, and BN(·) represents the batch normalization operation; the pixel rearrangement operation restores high-resolution details, and feature addition fuses global attention with local details, significantly improving the clarity of the edge of the segmentation result.

[0110] In this step, the design of a parameter-free similarity attention unit significantly reduces the model's computational complexity and avoids overfitting, while maintaining the ability to adaptively focus on key image regions. The combination of boundary and texture feature extraction units enables the model to simultaneously capture both boundary and texture information in an image, improving feature richness and expressiveness. Feature concatenation, fusion, and optimization further enhance feature robustness and segmentation accuracy.

[0111] like Figure 3 As shown, in this embodiment S3, the enhanced features are optimized and the category prediction is performed in combination with the segmentation network, and the segmentation result of the remote sensing image is output; specifically, the following steps are included:

[0112] S31, for the enhanced feature φ + Initial optimization is performed through convolutional layers, batch normalization layers, and activation functions to improve feature expression capabilities;

[0113] S32. Based on the enhanced features after initial optimization, multi-step denoising is performed through a diffusion model to gradually restore detail information and obtain optimized enhanced features. Specifically, in this embodiment, the diffusion model adopts a U-Net structure, which includes an encoder-decoder module. The encoder extracts multi-scale features through downsampling, and the decoder combines jump connections to gradually restore the resolution. The diffusion model gradually refines the features through multi-step denoising. Compared with single-step post-processing, it can more accurately restore small targets and fuzzy boundaries in complex scenes. It will be deeply integrated with the traditional segmentation network and denoising is performed in the feature space rather than the pixel space. It not only avoids the interference of the original image noise on the segmentation, but also significantly improves the detail recovery ability through iterative optimization.

[0114] S33. Combined with the segmentation head, the optimized enhanced features are converted into pixel-level classification results.

[0115] Furthermore, the above segmentation network adopts a composite loss function L total ;

[0116] L total =λ1L CE +λ2L Dice +λ3L Boundary

[0117] Among them, λ1, λ2, λ3 are weight coefficients, L CE represents the cross entropy loss function, L Dice represents the Dice loss function, L Boundary Represents the boundary loss function; specifically, cross entropy loss optimizes classification accuracy, Dice loss alleviates category imbalance, and boundary loss enhances edge sharpness. During the training process, the three work together to improve the comprehensive performance of the model in complex scenarios and can more comprehensively consider the various requirements of segmentation tasks.

[0118] In this step, the introduction of a diffusion model enables the model to gradually remove noise and restore image detail, significantly improving the clarity and accuracy of segmentation results, particularly when processing complex remote sensing images. Furthermore, the design of the segmentation head enables the model to accurately convert the optimized enhanced features into pixel-level classification results, achieving precise segmentation of remote sensing images.

[0119] This embodiment provides a remote sensing image segmentation method that incorporates a parameter-free similarity attention mechanism. Initial features are obtained through a feature extraction and fusion optimization network. A parameter-free similarity attention unit is used to enhance sensitivity in key regions. Furthermore, boundary and texture feature extraction units are combined to capture richer detail information. Furthermore, a diffusion model is used for multi-step denoising, gradually restoring and optimizing detail information. Ultimately, a segmentation network outputs a highly accurate segmentation result.

[0120] This method not only significantly reduces computational complexity and the risk of overfitting, but also improves the ability to capture boundaries and details in complex scenes through the integration of multiple techniques, achieving efficient and accurate remote sensing image segmentation. The use of a composite loss function further ensures classification accuracy and edge sharpness in the presence of class imbalance, improving the model's robustness and adaptability in practical applications.

[0121] Example 2;

[0122] like Figure 4 As shown, this embodiment takes the building segmentation in high-resolution urban remote sensing images as a specific scenario, and elaborates on the implementation process of the technical solution in detail, focusing on optimizing the challenges of dense buildings, shadow occlusion and complex boundaries in urban environments.

[0123] The overall process is as follows:

[0124] 1) Data preprocessing and feature extraction;

[0125] The feature extraction and fusion optimization network is fed with an urban remote sensing image with a resolution of 0.5m / pixel. The image contains dense buildings, roads, vegetation, and shadowed areas. First, image preprocessing is performed, using a physical model-based radiometric correction method to reduce the interference of building shadows on feature extraction. The CLAHE algorithm is then applied to enhance the visibility of low-contrast areas. Pixel values ​​are then normalized to the range [0, 1] to eliminate illumination variations.

[0126] Subsequently, the Mamba model pre-trained on the Urban3D dataset was loaded to extract high-level semantic features from the preprocessed images. Convolutional kernels of different sizes, 3×3, 5×7, and 7×7, were used to extract local window frame details, building outlines, and global block layout, respectively, to generate multi-level features. Through the weighted summation of channel attention, comprehensive features were generated, preserving the global structure and local details of the building. Finally, a 1×1 convolution was used to reduce the number of channels from 512 to 256 to generate the initial features.

[0127] 2) Feature enhancement and detail restoration;

[0128] In the feature enhancement stage, parameter-free similarity attention enhancement is first performed; the features are layer-normalized, and the mean and variance of each feature point after excluding itself are calculated to quantify its difference from surrounding features; a weight matrix is ​​generated through an energy function, and after Sigmoid adjustment, it is multiplied with the normalized features to obtain attention enhancement features, focusing on strengthening the edge areas of buildings.

[0129] Next, a two-way feature extraction and fusion method is adopted. A 3×3 dilated convolution with a dilation rate of r=2 is used to expand the receptive field to 7×7 to extract the building outline. At the same time, a depthwise separable convolution (3×3 channel-by-channel convolution + 1×1 point-by-point convolution) is used to extract the building surface texture. After splicing the two, a convolution+BN+ReLU operation and pixel rearrangement are performed to restore the high-resolution features and further sharpen the building edges.

[0130] 3) Diffusion model optimization and segmentation prediction;

[0131] During the diffusion model optimization phase, noise was initialized with step t = 100 based on the diffusion model, followed by a 50-step rapid denoising process. A lightweight U-Net architecture was employed, with the encoder performing 4-layer downsampling and the decoder performing 4-layer upsampling. Multi-scale features were fused via skip connections. During the denoising process, broken building outlines caused by shadows were effectively repaired and window frame textures were enhanced.

[0132] Furthermore, two layers of 3×3 convolution + 1×1 convolution are used to map the optimized features into pixel-level classification results. In the loss calculation, a composite loss function is used, including cross entropy loss (weight = 0.5), Dice loss (weight = 0.3) and boundary loss (weight = 0.2) to jointly optimize segmentation accuracy and edge consistency.

[0133] 4) post-processing;

[0134] Finally, post-processing steps are performed, including filling small holes, smoothing edge jaggedness, converting the segmentation results into vector polygons, and removing abnormal debris to obtain the final building segmentation results.

[0135] This embodiment focuses on building segmentation in high-resolution urban remote sensing images. Through data preprocessing, feature extraction, feature enhancement and detail recovery, diffusion model optimization and segmentation prediction, and post-processing steps, it effectively responds to challenges such as dense buildings, shadow occlusion, and complex boundaries, and ultimately obtains accurate building segmentation results.

[0136] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. References to the same or similar parts between the various embodiments are sufficient. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple. For relevant parts, refer to the method description.

[0137] 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 combining similarity attention without additional parameters, characterized in that The following steps are involved: S1. Using feature extraction and fusion optimization network, extract multi-scale features of remote sensing images and perform fusion optimization to generate initial features; S2. Based on the initial features, sensitivity enhancement is performed by the parameter-free similarity attention unit in the feature enhancement network, and boundary and texture feature extraction unit is used to extract boundary and texture features, and the enhanced features are generated by fusing them; The sensitivity enhancement comprises: S21, for the initial feature φ 0 Perform normalization to obtain normalized features S22, based on normalized features Calculate the mean and variance of each neuron to assess its relative difference from other neurons; where a neuron represents the eigenvalue corresponding to any position or channel; S23, through the energy function Quantify the neuron difference and generate the energy matrix E; the energy function Expressed as: Among them, t represents the current neuron, represents the mean of the remaining neurons after removing the current neuron, represents the variance of the remaining neurons after removing the current neuron, and λ represents the regularization parameter; S24, calculate the neuron importance weight according to the energy matrix E, and dynamically adjust the feature response through the Sigmoid function; S25. Multiply the adjusted features with the original features to generate attention-enhanced features Where sigmoid(·) represents the sigmoid activation function; S3. Combined with the segmentation network, the enhanced features are optimized and the category prediction is performed, and the segmentation result of the remote sensing image is output.

2. A remote sensing image segmentation method combining similarity attention without additional parameters according to claim 1, characterized in that: Said S1 comprises: S11. Use the Mamba pre-trained model to extract features from remote sensing images and obtain high-level semantic features φ; S12. Based on the high-level semantic feature φ, multi-scale feature extraction is performed through the multi-scale convolution layer to obtain a multi-level feature φ containing global information. multi ; S13, for multi-level features φ multi Perform weighted summation or concatenation fusion to generate comprehensive features φ fused ; S14, through the convolution layer and activation function to the comprehensive feature φ fused Perform dimensionality reduction optimization to obtain the initial feature φ 0 .

3. The remote sensing image segmentation method combined with similarity attention without additional parameters according to claim 2, characterized in that: In S12, the multi-scale convolution layer includes at least three convolution kernels of different sizes, and the multi-level feature φ multi Calculated by the following formula: φ multi =[Conv k1 ,Conv k2 ,Conv k2 ] Among them, Conv k1 、Conv k2 、Conv k2 They represent convolution operations with kernel sizes of k1, k2, and k3, respectively, and [·] represents a concatenation operation.

4. The remote sensing image segmentation method combined with similarity attention without additional parameters according to claim 1, characterized in that In S2, the boundary and texture feature extraction unit performs boundary and texture feature extraction, including: The initial feature φ 0 Input the hole convolution layer to expand the receptive field, and generate the boundary feature φ through the LeakyReLU activation function, convolution layer and ReLU activation function D ; φ D =ReLU(Conv(LeakyReLU(DConv(φ 0 )))) Among them, Conv(·) represents the convolutional layer operation, and DConv(·) represents the dilated convolutional layer operation; And, the initial feature φ 0 Input the depth separable convolution layer and process it through the Mish activation function and the fully connected layer to generate the texture feature φ DW ; f DW =FC(Mish(DWConv(φ 0 ))) Among them, FC(·) represents the fully connected layer operation, and DWConv(·) represents the depthwise separable convolutional layer operation.

5. The remote sensing image segmentation method combined with similarity attention without additional parameters according to claim 4, characterized in that: Also includes: The boundary feature φ D and texture features φ DW Splicing to obtain comprehensive features φ DDW ; For comprehensive feature φ DDW Convolution, batch normalization and ReLU activation are performed in sequence, and the boundary feature φ D and texture features φ DW Fusion to obtain enhanced features 6. The remote sensing image segmentation method combined with similarity attention without additional parameters according to claim 5, characterized in that: Said S2 further includes: Enhance the features Processed by ReLU activation function and enhanced with attention features after convolution and batch normalization Perform fusion to obtain fusion features; The fusion feature is pixel-rearranged to obtain the enhanced feature φ + ; Among them, PShuffle(·) represents the pixel rearrangement operation, and BN(·) represents the batch normalization operation.

7. The remote sensing image segmentation method combined with similarity attention without additional parameters according to claim 1, characterized in that: The S3 includes: S31, for the enhanced feature φ + Initial optimization is performed through convolutional layers, batch normalization layers, and activation functions to improve feature expression capabilities; S32, based on the enhanced features after initial optimization, performing multi-step denoising processing through a diffusion model to gradually restore detail information and obtain optimized enhanced features; S33. Combined with the segmentation head, the optimized enhanced features are converted into pixel-level classification results.

8. The remote sensing image segmentation method combined with similarity attention without additional parameters according to claim 7, characterized in that: The segmentation network adopts a composite loss function L total ; L total =λ1L CE +λ2L Dice +λ3L Boundary Among them, λ1, λ2, λ3 are weight coefficients, L CE represents the cross entropy loss function, L Dice represents the Dice loss function, L Boundary represents the boundary loss function.

Images