Remote sensing image segmentation method based on intensity grouping Transform network

Abstract

The invention discloses a remote sensing image segmentation method based on an intensity grouping Transform network, and relates to the field of remote sensing image segmentation, and the method comprises the following steps: carrying out the processing of a multi-stage preprocessing module, a convolution layer and an intensity grouping Transform module group on a remote sensing image through employing an encoder network, and generating features F1, F2, F3 and F4; the features F1, F2, F3 and F4 are subjected to multi-level convolutional layer and strength grouping Transform module group processing through a decoder network, and an initial segmentation result is generated; and performing post-processing optimization on the initial segmentation result, and outputting a segmentation result corresponding to the remote sensing image. According to the method, CNN local features and Transform global modeling are fused, and the remote sensing image segmentation precision, the boundary definition and the small target recognition capability are remarkably improved while the calculation overhead is reduced through adaptive grouping attention, pixel recombination, a shuffling technology and a multi-stage optimization strategy.

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 based on an intensity grouping Transformer network. Background Art

[0002] Remote sensing image segmentation is a key technology in fields such as land use classification, environmental monitoring, and disaster assessment. Its goal is to accurately classify different land features from high-resolution remote sensing imagery. However, remote sensing images have complex spectral information, large variations in land feature scales, and significant background noise, placing extremely high demands on the robustness and accuracy of segmentation algorithms.

[0003] Traditional convolutional neural networks (CNNs) rely on local receptive fields to extract features, making it difficult to effectively model long-range dependencies in remote sensing images, such as the relationship between large areas of water and surrounding vegetation. This limits segmentation accuracy in complex scenes. While the standard Transformer captures global relationships through a self-attention mechanism, its computational complexity increases quadratically with image resolution, making it difficult to directly apply to high-resolution remote sensing image processing and lacks targeted optimization for local details.

[0004] During the decoding phase, traditional upsampling methods (such as bilinear interpolation and deconvolution) require increasing the channel dimension to restore high-resolution features, which can lead to channel redundancy and a surge in computational overhead. Furthermore, these methods are prone to losing local details during reconstruction, causing blurred boundaries or artifacts, and are particularly incapable of capturing the details of small objects. While existing Transformer-based decoders attempt to improve global consistency, they fail to effectively incorporate efficient strategies such as channel reordering, making it difficult to strike a balance between computational efficiency and feature fidelity.

[0005] Existing methods often simply concatenate features from different levels when fusing multi-scale features, failing to fully incorporate mechanisms like adaptive sorting and grouped attention, leading to classification confusion for heterogeneous objects. Furthermore, post-processing often relies on single optimization techniques (such as morphological operations) without systematically enhancing boundary continuity and small object integrity, making it difficult to cope with noise interference and scale changes in complex scenes.

[0006] Therefore, how to design a remote sensing image segmentation method based on the intensity grouping Transformer network that can achieve efficient modeling of local details and global semantics, improve the segmentation accuracy in complex scenes through multi-level optimization strategies, and ensure the clarity of ground object boundaries and the ability to recognize small targets is an urgent problem that technical personnel 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 based on intensity grouping Transformer network. By integrating the local feature extraction capability of convolutional neural network with the global context modeling advantage of Transformer, and combining adaptive grouping attention mechanism, pixel reorganization, shuffling technology and multi-level optimization strategy, it solves the problems of difficulty in balancing local details and global features, high computational complexity, blurred boundaries and missed small targets in high-resolution remote sensing image segmentation.

[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 an intensity grouping Transformer network comprises the following steps:

[0010] S1, using the encoder network, processes the remote sensing image through a multi-level preprocessing module, convolutional layer and intensity grouping Transformer module group to generate features F1, F2, F3, and F4;

[0011] S2, through the decoder network, the features F1, F2, F3, and F4 are processed by multi-level convolutional layers and intensity grouping Transformer modules to generate the initial segmentation results;

[0012] S3. Perform post-processing optimization on the initial segmentation result and output a segmentation result corresponding to the remote sensing image.

[0013] Preferably, the S1 includes:

[0014] S11, preprocessing the remote sensing image through the first preprocessing module, the second preprocessing module, and the third preprocessing module in sequence to generate features f1, f2, and f3 respectively; and passing the remote sensing image through a 3×3 convolutional layer and a first intensity grouping Transformer module group to generate feature F1;

[0015] S12, reorganize the pixels of feature F1 and perform feature splicing with feature f1 to generate feature And the features The feature F2 is generated by processing it through a 1×1 convolutional layer and a second intensity grouping Transformer module group;

[0016] S13, reorganize the pixels of feature F2 and perform feature splicing with feature f2 to generate feature And the features The feature F3 is generated by processing it through a 1×1 convolutional layer and the third intensity grouping Transformer module group;

[0017] S14, reorganize the pixels of feature F3 and perform feature splicing with feature f3 to generate feature And the features The feature F4 is generated by processing it through a 1×1 convolutional layer and the fourth intensity grouping Transformer module group.

[0018] Preferably, the first preprocessing module, the second preprocessing module and the third preprocessing module all include: an average pooling layer, a 1×1 convolution layer and a 3×3 depthwise separable convolution layer.

[0019] Preferably, the pixel reorganization includes: dividing the features into 2×2 non-overlapping blocks, flattening each block into 4 channel vectors and splicing to the channel dimension, increasing the number of channels from C to 4C, and reducing the spatial size from H×W to H / 2×W / 2.

[0020] Preferably, the S2 includes:

[0021] S21, input feature F4 into the fifth intensity grouping Transformer module group, and perform pixel shuffling to generate feature 4 ′ F ;

[0022] S22, feature 4 ′ F Concatenate the feature with feature F3, process it through a 1×1 convolution layer and the sixth intensity grouping Transformer module group, and perform pixel shuffling to generate feature 3 ′ F ;

[0023] S23, feature 3 ′ F Concatenate the feature with feature F2, process it through a 1×1 convolution layer and the seventh intensity grouping Transformer module group, and perform pixel shuffling to generate feature 2 ′ F ;

[0024] S24, feature 2 ′ F The feature is concatenated with feature F1, and the initial segmentation result is generated through a 1×1 convolution layer, the eighth intensity grouping Transformer module group, a 3×3 convolution layer and an activation function.

[0025] Preferably, the pixel shuffling comprises: reducing the number of low-resolution feature channels C by r 2 Each group of channels is rearranged into r×r spatial blocks, which increases the spatial resolution of the output feature by r times and reduces the number of channels to C / r. 2 , where r is the preset upsampling ratio.

[0026] Preferably, each intensity grouping Transformer module group includes a plurality of intensity grouping Transformer modules, wherein the feature processing process of each intensity grouping Transformer module includes:

[0027] Pass the feature F′ through the normalization layer and perform channel segmentation to generate feature F 1' With F 2′ ;

[0028] For feature F 1' Perform the first sorting to generate indexes d1 and d2 respectively, and compare them with feature F 2′ Perform splicing to obtain feature F;

[0029] Input the feature F into the 1×1 convolution layer and the 3×3 depth-separable convolution layer, perform channel segmentation, and reshape them to obtain the feature F. a 、F b and F c ;

[0030] For feature F b Perform the second sort to generate index tensor d, feature F a 、F c Based on the index tensor d, index extraction is performed separately, and channel segmentation is performed to obtain the feature Q a , K a and Q c , K c ;

[0031] Q a , K a After reshaping, multiply and get weight W1, and Q c , K c After reshaping, multiply and obtain weight W2;

[0032] After sorting, feature F b Reshape and multiply by weight W1 and weight W2 respectively to obtain And After reshaping, perform point-wise multiplication to obtain F W ;

[0033] F W Perform channel segmentation to obtain features and Pair Features Perform a walk-through process and combine indexes d1 and d2 to generate recovery features and features Splicing is performed and the feature F is output through the 1*1 convolution layer + .

[0034] Preferably, the first sorting is vertical sorting and horizontal sorting in sequence, and the second sorting is vertical sorting or horizontal sorting.

[0035] Preferably, the S3 includes:

[0036] S31. Use conditional random fields to optimize the boundary continuity of the initial segmentation results and enhance the edge clarity of the objects;

[0037] S32, identify small target areas through connected component analysis and perform morphological dilation operation to enhance integrity;

[0038] S33. Use multi-scale Gaussian pyramid to perform feature fusion, extract detail information at different scales and reconstruct it by weight, and output the segmentation result corresponding to the remote sensing image.

[0039] Preferably, the intensity grouped Transformer network uses a composite loss function:

[0040] L total =λ1L CE +λ2L Dice

[0041] Among them, λ1 and λ2 represent weight coefficients, L CE represents the cross entropy loss function, L Dice represents the Dice loss function.

[0042] 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 advantages:

[0043] Beneficial effects:

[0044] 1. This method designs an intensity-grouped Transformer module, combining the local feature extraction advantages of convolutional neural networks with the global self-attention mechanism of Transformer to achieve efficient modeling of multi-scale features of remote sensing images. This module enhances the ability to capture the boundaries, textures, and spectral differences of heterogeneous objects through adaptive channel grouping, sorting, and group attention mechanisms, effectively improving segmentation accuracy in complex scenes.

[0045] 2. Pixel reorganization and pixel shuffling techniques are employed. During the encoding phase, spatial reorganization reduces resolution while increasing channel information, preserving local details. During the decoding phase, high-resolution restoration is achieved through channel rearrangement, avoiding edge blurring and artifacts caused by traditional upsampling methods. This strategy significantly reduces computational redundancy, ensures clear object boundaries, and enhances the ability to depict details in small target areas.

[0046] 3. Through a multi-level feature fusion mechanism, the multi-scale features extracted by the encoder and the high-resolution features recovered by the decoder are progressively spliced and integrated. A composite loss function, combining cross-entropy and Dice loss, balances pixel-level classification accuracy and regional similarity. Furthermore, conditional random fields (CRFs) and connected region analysis are introduced in the post-processing stage to further optimize boundary continuity and small object integrity, ensuring the robustness and reliability of segmentation results in complex remote sensing scenarios. 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 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.

[0048] Figure 1 A flow chart of a remote sensing image segmentation method based on an intensity grouping Transformer network provided by an embodiment of the present invention;

[0049] Figure 2 A diagram showing the structure of the intensity grouping Transformer network provided by an embodiment of the present invention;

[0050] Figure 3 This is a structural framework diagram of the intensity grouping Transformer module provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0051] 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.

[0052] like Figure 1 and Figure 2 As shown, this embodiment provides a remote sensing image segmentation method based on an intensity grouping Transformer network. The intensity grouping Transformer network includes an encoder network and a decoder network, and includes the following steps:

[0053] S1, using the encoder network, processes the remote sensing image through a multi-level preprocessing module, convolutional layer and intensity grouping Transformer module group to generate features F1, F2, F3, and F4;

[0054] S2, through the decoder network, the features F1, F2, F3, and F4 are processed by multi-level convolutional layers and intensity grouping Transformer modules to generate the initial segmentation results;

[0055] S3. Perform post-processing optimization on the initial segmentation result and output a segmentation result corresponding to the remote sensing image.

[0056] This method combines CNN local features with Transformer global modeling. Through adaptive group attention, pixel reorganization, shuffling and multi-level optimization strategies, it significantly improves remote sensing image segmentation accuracy, boundary clarity and small target recognition capabilities while reducing computational overhead.

[0057] The following further describes each step in the above method in detail;

[0058] In this embodiment S1, the encoder network is used to process the remote sensing image through a multi-level preprocessing module, a convolutional layer, and an intensity grouping Transformer module group to generate features F1, F2, F3, and F4; specifically, the following steps are included:

[0059] S11, preprocessing the remote sensing image through the first preprocessing module, the second preprocessing module, and the third preprocessing module in sequence to generate features f1, f2, and f3 respectively; and passing the remote sensing image through a 3×3 convolutional layer and a first intensity grouping Transformer module group to generate feature F1;

[0060] S12, reorganize the pixels of feature F1 and perform feature splicing with feature f1 to generate feature And the features The feature F2 is generated by processing it through a 1×1 convolutional layer and a second intensity grouping Transformer module group;

[0061] S13, reorganize the pixels of feature F2 and perform feature splicing with feature f2 to generate feature And the features The feature F3 is generated by processing it through a 1×1 convolutional layer and the third intensity grouping Transformer module group;

[0062] S14, reorganize the pixels of feature F3 and perform feature splicing with feature f3 to generate feature And the features The feature F4 is generated by processing it through a 1×1 convolutional layer and the fourth intensity grouping Transformer module group.

[0063] Furthermore, the first preprocessing module, the second preprocessing module, and the third preprocessing module all include: an average pooling layer, a 1×1 convolution layer, and a 3×3 depth-separable convolution layer; specifically, the average pooling layer is used for noise reduction and smoothing; the 1×1 convolution layer is used for channel dimension adjustment; and the 3×3 depth-separable convolution layer is used for efficient extraction of local features.

[0064] Furthermore, pixel reorganization includes: dividing the features into 2×2 non-overlapping blocks, flattening each block into 4 channel vectors and splicing them to the channel dimension, increasing the number of channels from C to 4C, and reducing the spatial size from H×W to H / 2×W / 2; the specific implementation of pixel reorganization here is:

[0065] The input feature map is divided into 2×2 non-overlapping blocks, and the four pixel values in each block (such as coordinates (i,j), (i,j+1), (i+1,j), (i+1,j+1)) are flattened into four channel vectors and concatenated according to the channel dimension; for example, the input size H×W×C becomes H / 2×W / 2×4C after reorganization, the spatial resolution is reduced but the channel information is doubled; it reduces the computational complexity of subsequent Transformer modules through spatial compression, and the increase in the number of channels enables the network to more finely characterize local spectral differences (such as the reflectance of different vegetation types).

[0066] In this step, the encoder reduces the number of parameters through depthwise separable convolution and combines it with the adaptive sorting mechanism of the intensity grouping Transformer to solve the computational bottleneck of the traditional Transformer in high-resolution remote sensing images. At the same time, it improves the ability to distinguish heterogeneous objects and realizes efficient multi-scale feature extraction, providing rich semantic and detail information for the decoding stage.

[0067] In this embodiment S2, features F1, F2, F3, and F4 are processed by a decoder network using a multi-level convolutional layer and an intensity grouping Transformer module group to generate an initial segmentation result. Specifically, the following steps are performed:

[0068] S21, input feature F4 into the fifth intensity grouping Transformer module group, and perform pixel shuffling to generate feature 4 ′ F ;

[0069] S22, feature 4 ′ F Concatenate the feature with feature F3, process it through a 1×1 convolution layer and the sixth intensity grouping Transformer module group, and perform pixel shuffling to generate feature 3 ′ F ;

[0070] S23, feature 3 ′ FConcatenate the feature with feature F2, process it through a 1×1 convolution layer and the seventh intensity grouping Transformer module group, and perform pixel shuffling to generate feature 2 ′ F ;

[0071] S24, feature 2 ′ F The feature is concatenated with feature F1, and the initial segmentation result is generated through a 1×1 convolution layer, the eighth intensity grouping Transformer module group, a 3×3 convolution layer and an activation function.

[0072] Furthermore, pixel shuffling includes: reducing the number of low-resolution feature channels C by r 2 Each group of channels is rearranged into r×r spatial blocks, which increases the spatial resolution of the output feature by r times and reduces the number of channels to C / r. 2 , where r is the preset upsampling ratio;

[0073] Assuming the input feature size is H×W×C and the upsampling ratio is r=2, the number of channels C is reduced by r 2 = 4 groups, with each group of C / 4 channels rearranged into a 2×2 spatial block, resulting in an output size of 2H×2W×C / 4. For example, the input channels [1, 2, 3, 4] are rearranged into a 2×2 spatial block, restoring high-resolution details. This channel rearrangement avoids the checkerboard effect of deconvolution, making boundaries clearer, and reduces computational effort by reducing channel dimensionality, thereby improving resolution.

[0074] Furthermore, each intensity-grouped Transformer module group includes multiple intensity-grouped Transformer modules. Specifically, in the encoder network, the first intensity-grouped Transformer module group includes 4 intensity-grouped Transformer modules connected in series, the second intensity-grouped Transformer module group includes 4 intensity-grouped Transformer modules connected in series, the third intensity-grouped Transformer module group includes 6 intensity-grouped Transformer modules connected in series, and the third intensity-grouped Transformer module group includes 8 intensity-grouped Transformer modules connected in series; in the decoder network, the fifth intensity-grouped Transformer module group includes 8 intensity-grouped Transformer modules connected in series, the sixth intensity-grouped Transformer module group includes 6 intensity-grouped Transformer modules connected in series, the seventh intensity-grouped Transformer module group includes 4 intensity-grouped Transformer modules connected in series, and the eighth intensity-grouped Transformer module group includes 4 intensity-grouped Transformer modules connected in series;

[0075] like Figure 3 As shown, the feature processing process of each intensity grouping Transformer module includes:

[0076] Pass the feature F′ through the normalization layer and perform channel segmentation to generate feature F 1' With F 2′ ;

[0077] For feature F 1′ Perform the first sorting to generate indexes d1 and d2 respectively, and compare them with feature F 2′ Perform splicing to obtain feature F;

[0078] Input the feature F into the 1×1 convolution layer and the 3×3 depth-separable convolution layer, perform channel segmentation, and reshape them to obtain the feature F. a 、F b and F c ;

[0079] For feature F b Perform the second sort to generate index tensor d, feature F a 、F c Based on the index tensor d, index extraction is performed separately, and channel segmentation is performed to obtain the feature Q a , K a and Q c , K c ;

[0080] Q a , K a After reshaping, multiply and get weight W1, and Q c , K c After reshaping, multiply and obtain weight W2;

[0081] After sorting, feature F b Reshape and multiply by weight W1 and weight W2 respectively to obtain And After reshaping, perform point-wise multiplication to obtain F W ;

[0082] F W Perform channel segmentation to obtain features and Pair Features Perform a walk-through process and combine indexes d1 and d2 to generate recovery features and features Splicing is performed and feature F is output through a 1*1 convolutional layer + .

[0083] Furthermore, the above-mentioned first sorting is vertical sorting and horizontal sorting in sequence, and the second sorting is vertical sorting or horizontal sorting; the vertical sorting and horizontal sorting specifically include: vertical sorting is to arrange the feature columns in ascending or descending order according to pixel intensity along the height direction to generate a vertical sorting index; horizontal sorting is to sort the feature rows by pixel intensity twice along the width direction on the basis of vertical sorting to generate a horizontal sorting index.

[0084] This module uses dynamic channel segmentation and dual sorting mechanisms to adaptively focus on high-difference areas, enhancing the ability to distinguish spectral and spatial features. It combines group attention calculation and depth-wise separable convolution to efficiently model global context and local details while reducing computational complexity. It ensures lossless information transmission through index restoration and feature fusion, significantly improving the segmentation accuracy and robustness of complex scenes (such as mixed urban areas and small target dense areas).

[0085] In this embodiment S3, post-processing and optimizing the initial segmentation result is performed to output a segmentation result corresponding to the remote sensing image; specifically, the following steps are performed:

[0086] S31. Use conditional random fields to optimize the boundary continuity of the initial segmentation results and enhance the edge clarity of the objects;

[0087] S32, identify small target areas through connected component analysis and perform morphological dilation operation to enhance integrity;

[0088] S33. Use multi-scale Gaussian pyramids for feature fusion, extract detail information at different scales and reconstruct it by weight, and output the segmentation results corresponding to the remote sensing image; here, feature representations of different scales are constructed through Gaussian pyramids: the initial segmentation results are subjected to multi-level Gaussian filtering and downsampling to generate multi-resolution feature maps from coarse to fine, and then the detail information at different scales is integrated through weighted fusion, thereby enhancing the model's adaptability to large and small targets.

[0089] Post-processing optimization optimizes boundary continuity through conditional random fields, identifies and enhances small targets through connected region analysis, and weightedly integrates features of different resolutions through multi-scale pyramid fusion, ultimately improving the boundary clarity, small target integrity and scale robustness of the segmentation results.

[0090] Furthermore, the above intensity grouping Transformer network adopts a composite loss function:

[0091] L total =λ1L CE +λ2L Dice

[0092] Among them, λ1 and λ2 represent weight coefficients, L CE represents the cross entropy loss function, L Dicerepresents the Dice loss function.

[0093] The composite loss function enhances pixel-level classification accuracy through cross-entropy loss to ensure accurate distinction of ground object categories; further combined with Dice loss to optimize target area matching, alleviate the common problems of missed detection of small targets and category imbalance in remote sensing images; and balances local details and global consistency by dynamically adjusting weight coefficients, for example, improving the recall rate of small targets in complex scenes, ultimately significantly improving the model's intersection-over-union ratio and achieving high-precision and robust remote sensing image segmentation.

[0094] The remote sensing image segmentation method based on the intensity-grouped Transformer proposed in this example achieves a deep fusion of local features and global context through an encoder-decoder architecture. The encoder uses pixel reorganization and adaptive sorting to reduce computational complexity while preserving multi-scale details. The decoder uses pixel shuffling to efficiently recover high-resolution features and combines grouped attention to optimize boundary accuracy. The post-processing strategy further improves the reliability of the results through CRF and multi-scale fusion.

[0095] 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.

[0096] 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 intensity grouping Transformer network, characterized in that: The following steps are involved: S1, using the encoder network, processes the remote sensing image through a multi-level preprocessing module, convolutional layer and intensity grouping Transformer module group to generate features F1, F2, F3, and F4; S2, through the decoder network, the features F1, F2, F3, and F4 are processed by multi-level convolutional layers and intensity grouping Transformer modules to generate the initial segmentation results; S3. Perform post-processing optimization on the initial segmentation result and output a segmentation result corresponding to the remote sensing image.

2. The remote sensing image segmentation method based on intensity grouping Transformer network according to claim 1, characterized in that: Said S1 comprises: S11, preprocessing the remote sensing image through the first preprocessing module, the second preprocessing module, and the third preprocessing module in sequence to generate features f1, f2, and f3 respectively; and passing the remote sensing image through a 3×3 convolutional layer and a first intensity grouping Transformer module group to generate feature F1; S12, reorganize the pixels of feature F1 and perform feature splicing with feature f1 to generate feature And the features The feature F2 is generated by processing it through a 1×1 convolutional layer and a second intensity grouping Transformer module group; S13, reorganize the pixels of feature F2 and perform feature splicing with feature f2 to generate feature And the features The feature F3 is generated by processing it through a 1×1 convolutional layer and the third intensity grouping Transformer module group; S14, reorganize the pixels of feature F3 and perform feature splicing with feature f3 to generate feature And the features The feature F4 is generated by processing it through a 1×1 convolutional layer and the fourth intensity grouping Transformer module group.

3. The remote sensing image segmentation method based on intensity grouping Transformer network according to claim 2, characterized in that: The first preprocessing module, the second preprocessing module, and the third preprocessing module all include: an average pooling layer, a 1×1 convolution layer, and a 3×3 depth-separable convolution layer.

4. The remote sensing image segmentation method based on intensity grouping Transformer network according to claim 2, characterized in that: The pixel reorganization includes: dividing the features into 2×2 non-overlapping blocks, flattening each block into 4 channel vectors and splicing to the channel dimension, increasing the number of channels from C to 4C, and reducing the spatial size from H×W to H / 2×W / 2.

5. The remote sensing image segmentation method based on intensity grouping Transformer network according to claim 1, characterized in that: The S2 includes: S21, input feature F4 into the fifth intensity grouping Transformer module group, and perform pixel shuffling to generate feature 4 ′ F ; S22, feature 4 ′ F Concatenate the feature with feature F3, process it through a 1×1 convolution layer and the sixth intensity grouping Transformer module group, and perform pixel shuffling to generate feature 3 ′ F ; S23, feature 3 ′ F Concatenate the feature with feature F2, process it through a 1×1 convolution layer and the seventh intensity grouping Transformer module group, and perform pixel shuffling to generate feature 2 ′ F ; S24, feature 2 ′ F The feature is concatenated with feature F1, and the initial segmentation result is generated through a 1×1 convolution layer, the eighth intensity grouping Transformer module group, a 3×3 convolution layer and an activation function.

6. The remote sensing image segmentation method based on intensity grouping Transformer network according to claim 5, characterized in that: The pixel shuffling includes: reducing the number of low-resolution feature channels C by r 2 Each group of channels is rearranged into r×r spatial blocks, which increases the spatial resolution of the output feature by r times and reduces the number of channels to C / r. 2 , where r is the preset upsampling ratio.

7. The remote sensing image segmentation method based on intensity grouping Transformer network according to claim 1, characterized in that: Each intensity-grouped Transformer module group includes multiple intensity-grouped Transformer modules, wherein the feature processing process of each intensity-grouped Transformer module includes: Pass the feature F′ through the normalization layer and perform channel segmentation to generate feature F 1' With F 2′ ; For feature F 1′ Perform the first sorting to generate indexes d1 and d2 respectively, and compare them with feature F 2′ Perform splicing to obtain feature F; Input the feature F into the 1×1 convolution layer and the 3×3 depth-separable convolution layer, perform channel segmentation, and reshape them to obtain the feature F. a 、F b and F c ; For feature F b Perform the second sort to generate index tensor d, feature F a 、F c Based on the index tensor d, index extraction is performed separately, and channel segmentation is performed to obtain the feature Q a , K a and Q c , K c ; Q a , K a After reshaping, multiply and get weight W1, and Q c , K c After reshaping, multiply and obtain weight W2; After sorting, feature F b Reshape and multiply by weight W1 and weight W2 respectively to obtain And After reshaping, perform point-wise multiplication to obtain F W ; F W Perform channel segmentation to obtain features and Pair Features Perform a walk-through process and combine indexes d1 and d2 to generate recovery features and features Splicing is performed and the feature F is output through the 1*1 convolution layer + .

8. The remote sensing image segmentation method based on intensity grouping Transformer network according to claim 7, characterized in that: The first sorting is vertical sorting and horizontal sorting in sequence, and the second sorting is vertical sorting or horizontal sorting.

9. The remote sensing image segmentation method based on intensity grouping Transformer network according to claim 1, characterized in that: The S3 includes: S31. Use conditional random fields to optimize the boundary continuity of the initial segmentation results and enhance the edge clarity of the objects; S32, identify small target areas through connected component analysis and perform morphological dilation operation to enhance integrity; S33. Use multi-scale Gaussian pyramid to perform feature fusion, extract detail information at different scales and reconstruct it by weight, and output the segmentation result corresponding to the remote sensing image.

10. The remote sensing image segmentation method based on intensity grouping Transformer network according to claim 1, characterized in that: The intensity grouped Transformer network uses a composite loss function: L total =λ1L CE +λ2L Dice Among them, λ1 and λ2 represent weight coefficients, L CE represents the cross entropy loss function, L Dice represents the Dice loss function.

Images