A remote sensing image segmentation method based on an intensity grouping transformer network

Abstract

The application discloses a remote sensing image segmentation method based on an intensity grouping Transformer network, and relates to the field of remote sensing image segmentation, and comprises the following steps: using an encoder network, performing multi-level preprocessing module, convolution layer and intensity grouping Transformer module group processing on a remote sensing image to generate features F1, F2, F3 and F4; through a decoder network, performing multi-level convolution layer and intensity grouping Transformer module group processing on the features F1, F2, F3 and F4 to generate an initial segmentation result; performing post-processing optimization on the initial segmentation result to output a segmentation result corresponding to the remote sensing image. The method fuses CNN local features and Transformer global modeling, and through adaptive grouping attention, pixel reorganization, shuffling technology and multi-level optimization strategies, the calculation overhead is reduced, and the remote sensing image segmentation precision, boundary definition and small target recognition capability are significantly improved.

Description

TECHNICAL FIELD

[0001] The present application 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

[0002] Remote sensing image segmentation is a key technology in the fields of land use classification, environmental monitoring, disaster assessment, etc., and its goal is to accurately divide different ground object categories from high-resolution remote sensing images. However, remote sensing images have the characteristics of complex spectral information, large differences in ground object scales, and significant background noise, which puts high requirements on the robustness and accuracy of segmentation algorithms.

[0003] Traditional convolutional neural networks (CNNs) rely on local receptive fields to extract features, which makes it difficult to effectively model the global contextual information of long-range dependencies in remote sensing images (such as the relevance of large-area water bodies and surrounding vegetation), resulting in limited segmentation accuracy in complex scenarios. While standard Transformers capture global relationships through self-attention mechanisms, their computational complexity grows quadratically with image resolution, making them difficult to directly apply to high-resolution remote sensing image processing, and lacking targeted optimization for local details.

[0004] In the decoding stage, traditional upsampling methods (such as bilinear interpolation and deconvolution) need to restore high-resolution features by increasing the channel dimension, which can easily lead to channel redundancy and a sharp increase in computational overhead. In addition, such methods are prone to lose local details during reconstruction, causing boundary blurring or artifact problems, especially lacking the ability to capture details of small targets. Existing Transformers-based decoders attempt to improve global consistency, but fail to effectively combine efficient strategies such as channel rearrangement, making it difficult to balance computational efficiency and feature fidelity.

[0005] Moreover, existing methods often simply concatenate features at different levels when fusing multi-scale features, without fully utilizing adaptive ordering and grouping attention mechanisms, leading to confusion in the classification of heterogeneous objects. In addition, post-processing often relies on a single optimization method (such as morphological operations), lacking systematic enhancement of boundary continuity and small target integrity, making it difficult to deal with noise interference and scale changes in complex scenarios.

[0006] Therefore, how to design a remote sensing image segmentation method based on an intensity grouping Transformer network that can effectively model local details and global semantics, improve segmentation accuracy in complex scenarios through multi-level optimization strategies, and ensure ground object boundary clarity and small target recognition ability is a problem that needs to be solved by those skilled in the art. SUMMARY

[0007] Therefore, the application provides a remote sensing image segmentation method based on an intensity grouping Transformer network.

[0008] To achieve the above object, the application adopts the following technical scheme:

[0009] A remote sensing image segmentation method based on an intensity grouping Transformer network comprises the following steps:

[0010] S1, using an encoder network, performing multi-level preprocessing module, convolution layer and intensity grouping Transformer module group processing on the remote sensing image to generate features F1, F2, F3 and F4;

[0011] S2, using a decoder network, performing multi-level convolution layer and intensity grouping Transformer module group processing on the features F1, F2, F3 and F4 to generate an initial segmentation result;

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

[0013] Preferably, S1 comprises:

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

[0015] S12, performing pixel reorganization on the feature F1 and feature splicing with the feature f1 to generate a feature and processing the feature through a 1*1 convolution layer and a second intensity grouping Transformer module group to generate a feature F2;

[0016] S13, performing pixel reorganization on the feature F2 and feature splicing with the feature f2 to generate a feature and processing the feature through a 1*1 convolution layer and a third intensity grouping Transformer module group to generate a feature F3;

[0017] S14, performing pixel reorganization on the feature F3 and feature splicing with the feature f3 to generate a feature and the features The features F4 are generated by processing through a 1x1 convolutional layer and a fourth intensity grouping Transformer module group.

[0018] Preferably, the first preprocessing module, the second preprocessing module, and the third preprocessing module each include an average pooling layer, a 1x1 convolutional layer, and a 3x3 depthwise separable convolutional layer.

[0019] Preferably, the pixel reorganization includes dividing the features into 2x2 non-overlapping blocks, flattening each block into 4 channel vectors and concatenating to the channel dimension, increasing the number of channels from C to 4C, and reducing the spatial size from HxW to H / 2xW / 2.

[0020] Preferably, the S2 includes:

[0021] S21, inputting the features F4 into a fifth intensity grouping Transformer module group and performing pixel shuffling to generate features

[0022] S22, inputting the features into a 1x1 convolutional layer and a sixth intensity grouping Transformer module group, performing pixel shuffling to generate features

[0023] S23, inputting the features into a 1x1 convolutional layer and a seventh intensity grouping Transformer module group, performing pixel shuffling to generate features

[0024] S24, inputting the features into a 1x1 convolutional layer, an eighth intensity grouping Transformer module group, a 3x3 convolutional layer, and an activation function to generate an initial segmentation result.

[0025] Preferably, the pixel shuffling includes grouping the number of channels C of the low-resolution features by r 2 times, rearranging the channels in each group into rxr spatial blocks, and increasing the spatial resolution of the output features by r times and reducing the number of channels to C / r 2 , where r is a preset up-sampling rate.

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

[0027] inputting the features F 'Feature F is generated through a normalization layer and channel segmentation. 1' With F 2' ;

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

[0029] The feature F is input into a 1×1 convolutional layer and a 3×3 depthwise separable convolutional layer. After channel segmentation, the features are reshaped to obtain the new features F. a F b and F c ;

[0030] For feature F b Perform a second sort to generate an index tensor d, and features F. a f c Based on the index tensor d, index extraction is performed, and then channel segmentation is performed to obtain the feature Q. a K a And Q c K c ;

[0031] Q a K a After reshaping and multiplying, obtain the weight W1, and then Q... c K c After reshaping, multiply the results to obtain the weight W2;

[0032] The features F after sorting b Reshape the product by multiplying it by weights W1 and W2 respectively, to obtain... And on After reshaping, perform point-by-point multiplication to obtain F. W ;

[0033] F W Channel segmentation to obtain features and Features Perform scattering processing and combine indices d1 and d2 to generate restored features and features. The features F are concatenated and output through a 1x1 convolutional layer. + .

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

[0035] Preferably, S3 includes:

[0036] S31, the boundary continuity of the initial segmentation result is optimized by using a conditional random field, and the ground object edge definition is enhanced;

[0037] S32, small target regions are identified by connected region analysis, and a morphological dilation operation is performed to enhance completeness;

[0038] S33, feature fusion is performed using a multi-scale Gaussian pyramid, detail information at different scales is extracted and weighted reconstruction is performed, and a segmentation result corresponding to the remote sensing image is output.

[0039] Preferably, the intensity grouping Transformer network adopts a composite loss function:

[0040] L total =λ1L CE +λ2L Dice

[0041] Wherein, λ1, λ2 represents the weight coefficient, L CE represents the cross-entropy loss function, L Dice represents the Dice loss function.

[0042] According to the technical scheme, compared with the prior art, the technical scheme has the following beneficial effects:

[0043] 1. The method designs an intensity grouping Transformer module, combines the local feature extraction advantage of the convolutional neural network and the global self-attention mechanism of the Transformer, and realizes efficient modeling of the multi-scale features of the remote sensing image. The module enhances the capture ability of the heterogeneous ground object boundary, texture and spectral difference through adaptive channel grouping, sorting and grouping attention mechanism, and effectively improves the segmentation precision in complex scenes.

[0044] 2. Pixel recombination and pixel shuffling technology is adopted, the spatial recombination strategy is used in the encoding stage to reduce the resolution while increasing the channel information, and the local details are retained; the high resolution is restored by channel rearrangement in the decoding stage, avoiding the edge blur or artifact problem caused by the traditional upsampling method. This strategy significantly reduces the calculation redundancy, ensures the ground object boundary definition, and improves the small target region detail description ability.

[0045] 3. Through the multi-level feature fusion mechanism, the multi-scale features extracted by the encoder and the high resolution features recovered by the decoder are spliced and integrated level by level, and the composite loss function of cross-entropy and Dice loss is combined, which takes into account the pixel-level classification accuracy and regional similarity. In addition, the conditional random field (CRF) and connected region analysis are introduced in the post-processing stage to further optimize the boundary continuity and small target integrity, ensuring the robustness and reliability of the segmentation result in complex remote sensing scenes. BRIEF DESCRIPTION OF DRAWINGS

[0046] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the accompanying drawings needed to be used in the embodiments or prior art description will be briefly introduced as follows. Obviously, the accompanying drawings in the following description only only the embodiments of the present application, and for those skilled in the art, other drawings can be obtained without creative labor on the basis of the provided drawings.

[0047] Figure 1 A flow chart of a remote sensing image segmentation method based on an intensity grouping Transformer network is provided for the embodiments of the present application.

[0048] Figure 2 A structure framework diagram of an intensity grouping Transformer network is provided for the embodiments of the present application.

[0049] Figure 3 A structure framework diagram of an intensity grouping Transformer module is provided for the embodiments of the present application. DETAILED DESCRIPTION

[0050] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments only are some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of the present application.

[0051] As shown in Figure 1 and Figure 2 , the present embodiment provides a remote sensing image segmentation method based on an intensity grouping Transformer network, which includes an encoder network and a decoder network, and includes the following steps:

[0052] S1, using the encoder network, performing multi-level preprocessing module, convolution layer and intensity grouping Transformer module group processing on the remote sensing image to generate features F1, F2, F3 and F4;

[0053] S2, through the decoder network, performing multi-level convolution layer and intensity grouping Transformer module group processing on the features F1, F2, F3 and F4 to generate an initial segmentation result;

[0054] S3, performing post-processing optimization on the initial segmentation result to output a segmentation result corresponding to the remote sensing image.

[0055] The method fuses the local features of the CNN and the global modeling of the Transformer, and through adaptive grouping attention, pixel reorganization, shuffling and multi-level optimization strategies, the calculation overhead is reduced while the remote sensing image segmentation accuracy, boundary definition and small target recognition ability are significantly improved.

[0056] The following further details each step in the above method;

[0057] In this embodiment S1, the encoder network is used to process the remote sensing image through a multi-level preprocessing module, a convolution layer and a strength grouping Transformer module, to generate features F1, F2, F3 and F4; specifically including:

[0058] S11, the remote sensing image is sequentially preprocessed through the first preprocessing module, the second preprocessing module and the third preprocessing module to generate features f1, f2 and f3 respectively; and the remote sensing image is processed through a 3x3 convolution layer and a first strength grouping Transformer module to generate a feature F1;

[0059] S12, the feature F1 is pixel reorganized and spliced with the feature f1 to generate a feature and the feature is processed through a 1x1 convolution layer and a second strength grouping Transformer module to generate a feature F2;

[0060] S13, the feature F2 is pixel reorganized and spliced with the feature f2 to generate a feature and the feature is processed through a 1x1 convolution layer and a third strength grouping Transformer module to generate a feature F3;

[0061] S14, the feature F3 is pixel reorganized and spliced with the feature f3 to generate a feature and the feature is processed through a 1x1 convolution layer and a fourth strength grouping Transformer module to generate a feature F4.

[0062] Further, the first preprocessing module, the second preprocessing module and the third preprocessing module each include an average pooling layer, a 1x1 convolution layer and a 3x3 depth separable convolution layer; specifically, the average pooling layer is used for noise reduction and smoothing processing; the 1x1 convolution layer is used for channel dimension adjustment; and the 3x3 depth separable convolution layer is used for efficient extraction of local features.

[0063] Further, the pixel reorganization includes: dividing the features into 2x2 non-overlapping blocks, flattening each block into 4 channel vectors and splicing to the channel dimension, the number of channels increasing from C to 4C, and the spatial size decreasing from HxW to H / 2xW / 2; the specific implementation of the pixel reorganization is:

[0064] The input feature map is divided into 2x2 non-overlapping blocks, and the 4 pixel values (such as coordinates (i, j), (i, j+1), (i+1, j), (i+1, j+1)) in each block are flattened into 4 channel vectors and spliced in the channel dimension; for example, the input size HxWxC is reorganized into H / 2xW / 2x4C, the spatial resolution is reduced but the channel information is doubled; it reduces the computational amount of the subsequent Transformer module through spatial compression, and the increase in the number of channels enables the network to more finely represent local spectral differences (such as reflectivity of different vegetation types).

[0065] In this step, the encoder reduces the parameter amount through deep separable convolution, combines the adaptive ordering mechanism of the intensity grouping Transformer, solves the computational bottleneck of the traditional Transformer in high-resolution remote sensing images, improves the ability to distinguish heterogeneous objects, and realizes efficient multi-scale feature extraction, providing rich semantic and detail information for the decoding stage.

[0066] In this embodiment S2, the features F1, F2, F3 and F4 are processed by a multi-level convolutional layer and an intensity grouping Transformer module group through a decoder network to generate an initial segmentation result; specifically including:

[0067] S21, input the feature F4 into the fifth intensity grouping Transformer module group, and generate a feature F4' through pixel shuffling;

[0068] S22, input the feature F3 into the sixth intensity grouping Transformer module group, and generate a feature F3' through pixel shuffling;

[0069] S23, input the feature F2 into the seventh intensity grouping Transformer module group, and generate a feature F2' through pixel shuffling;

[0070] S24, input the feature F1 into the eighth intensity grouping Transformer module group, and generate an initial segmentation result through a 1x1 convolutional layer, a 3x3 convolutional layer and an activation function. ​​​

[0071] Further, the pixel shuffling includes: grouping the low-resolution feature channel number C by r 2 times, rearranging each group of channels into an r×r spatial block, so that the output feature spatial resolution is increased by r times, and the channel number is reduced to C / r 2 , wherein r is a preset up-sampling rate;

[0072] Assuming that the input feature size is H×W×C, and the up-sampling rate r=2, the channel number C is grouped by r 2 =4, and each group of C / 4 channels is rearranged into a 2×2 spatial block, and the output size becomes 2H×2W×C / 4. For example, the input channel [1, 2, 3, 4] is rearranged to form a 2×2 spatial block, and the high-resolution details are recovered; it avoids the checkerboard effect of deconvolution through channel rearrangement, makes the boundary clearer, and reduces the calculation amount by reducing the channel dimension, and improves the resolution.

[0073] Further, each intensity grouping Transformer module group includes a plurality of intensity grouping Transformer modules. Specifically, in the encoder network, the first intensity grouping Transformer module group includes 4 intensity grouping Transformer modules connected in series, the second intensity grouping Transformer module group includes 4 intensity grouping Transformer modules connected in series, the third intensity grouping Transformer module group includes 6 intensity grouping Transformer modules connected in series, and the third intensity grouping Transformer module group includes 8 intensity grouping Transformer modules connected in series; in the decoder network, the fifth intensity grouping Transformer module group includes 8 intensity grouping Transformer modules connected in series, the sixth intensity grouping Transformer module group includes 6 intensity grouping Transformer modules connected in series, the seventh intensity grouping Transformer module group includes 4 intensity grouping Transformer modules connected in series, and the eighth intensity grouping Transformer module group includes 4 intensity grouping Transformer modules connected in series.

[0074] As shown in Figure 3 , the feature processing process of each intensity grouping Transformer module includes:

[0075] normalizing the feature F ' through a normalization layer, and performing channel segmentation to generate a feature F 1' and F 2' ;

[0076] performing first sorting on the feature F 1' , respectively generating indexes d1 and d2, and combining the feature F2' The features F are input into a 1*1 convolution layer, a 3*3 depth separable convolution layer, and are respectively reshaped to obtain features F

[0077] The features F are input into a 1*1 convolution layer, a 3*3 depth separable convolution layer, and are respectively reshaped to obtain features F a , F b , and F c ;

[0078] The features F b are subjected to a second sorting to generate an index tensor d, and the features F a , F c are respectively indexed and subjected to channel segmentation to obtain features Q a , K a , and Q c , K c ;

[0079] The Q a , K a are multiplied after being reshaped to obtain a weight W1, and the Q c , K c are multiplied after being reshaped to obtain a weight W2;

[0080] The features F b after sorting are respectively multiplied by the weights W1 and W2 to obtain and the F W after being reshaped are point-by-point multiplied to obtain F ;

[0081] The F W are subjected to channel segmentation to obtain features and The features are subjected to scattering processing, and combined with the indexes d1 and d2 to generate a restored feature and a feature are spliced, and the features F + are output through a 1*1 convolution layer.

[0082] Further, the first sorting is sequentially performed vertical sorting and horizontal sorting, and the second sorting is vertical sorting or horizontal sorting; the vertical sorting and the horizontal sorting specifically include: the vertical sorting is arranging the feature columns along the height direction in ascending or descending order of pixel intensity to generate a vertical sorting index; and the horizontal sorting is performing secondary sorting on the feature rows along the width direction on the basis of the vertical sorting to generate a horizontal sorting index.

[0083] The module focuses on high difference areas adaptively through dynamic channel segmentation and double sorting mechanism, and enhances the distinguishing ability of spectral and spatial features; combined with grouping attention calculation and depth separable convolution, the global context and local details are modeled efficiently while reducing the computational complexity; through index restoration and feature fusion, information is transmitted losslessly, and the segmentation accuracy and robustness of complex scenes (such as urban mixed areas and small target dense areas) are significantly improved.

[0084] In the embodiment S3, the initial segmentation result is post-processed and optimized to output a segmentation result corresponding to the remote sensing image; specifically including:

[0085] S31, the boundary continuity of the initial segmentation result is optimized by using conditional random field, and the edge definition of the ground object is enhanced;

[0086] S32, small target areas are identified by connected region analysis, and morphological dilation operation is performed to enhance the integrity;

[0087] S33, feature fusion is performed by using multi-scale Gaussian pyramid to extract detailed information at different scales and weighted reconstruction, and a segmentation result corresponding to the remote sensing image is output; here, different scale feature representations are constructed by Gaussian pyramid: multi-level Gaussian filtering and down-sampling are performed on the initial segmentation result to generate multi-resolution feature maps from coarse to fine, and then the detailed information at different scales is integrated by weighted fusion, so that the adaptability of the model to large and small targets is enhanced.

[0088] The post-processing optimization optimizes the boundary continuity by conditional random field, identifies and enhances small targets by connected region analysis, and integrates different resolution features by multi-scale pyramid fusion and weighted integration, so as to finally improve the boundary definition, small target integrity and scale robustness of the segmentation result.

[0089] Further, the above intensity grouping Transformer network adopts a composite loss function:

[0090] L total =λ1L CE +λ2L Dice

[0091] Wherein, λ1, λ2 represent weight coefficients, L CE represents a cross-entropy loss function, and L Dice represents a Dice loss function.

[0092] The composite loss function enhances the pixel-level classification accuracy through cross-entropy loss to ensure accurate distinction of feature categories, and further combines the Dice loss to optimize the target region matching degree, thereby relieving the small target missing detection and category imbalance problems commonly existing in remote sensing images; the weight coefficient is dynamically adjusted to balance the local details and global consistency, for example, to improve the small target recall rate in complex scenes, so that the model intersection over union is significantly improved, and high-precision and robust remote sensing image segmentation is realized.

[0093] The remote sensing image segmentation method based on intensity grouping Transformer proposed in the embodiment realizes deep fusion of local features and global context through an encoder-decoder architecture. The encoder adopts a pixel reorganization and adaptive ordering mechanism to reduce the computational load while retaining multi-scale details; the decoder efficiently recovers high-resolution features by using pixel shuffling, and optimizes the boundary accuracy by combining grouping attention; the post-processing strategy further improves the result reliability through CRF and multi-scale fusion.

[0094] The embodiments in the specification are described in a progressive manner, and each embodiment focuses on the difference from other embodiments. The same or similar parts of each embodiment can be referred to each other. For the system disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple, and the related parts can be referred to the method part.

[0095] The above description of the disclosed embodiments enables a person skilled in the art to implement or use the present application. Various modifications to the embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A remote sensing image segmentation method based on an intensity grouping Transformer network, characterized in that, The method comprises the following steps: S1, using an encoder network, performing multi-level preprocessing module, convolution layer and intensity grouping Transformer module group processing on the remote sensing image to generate features F1, F2, F3 and F4; S2, through a decoder network, performing multi-level convolution layer and intensity grouping Transformer module group processing on the features F1, F2, F3 and F4 to generate an initial segmentation result; S3, performing post-processing optimization on the initial segmentation result to output a segmentation result corresponding to the remote sensing image; Each intensity grouping Transformer module group comprises a plurality of intensity grouping Transformer modules, wherein the feature processing process of each intensity grouping Transformer module comprises: The feature F ' By normalizing layer, and carrying out channel segmentation to generate the feature F 1' With F 2' ; For feature F 1' First sorting, respectively generating indexes d1 and d2, and with feature F 2' Carrying out splicing, obtaining feature F; The feature F is input into a 1x1 convolutional layer, a 3x3 depth separable convolutional layer, and after channel segmentation, reshaping is performed to obtain features F a , F b , and F c ; On the feature F b The second sorting generates an index tensor d, the feature F a , F c Based on the index tensor d, the index extraction is carried out respectively, and the channel segmentation is carried out to obtain the feature Q a , K a And Q c , K c ; Q a , K a are multiplied after reshaping to obtain weight W1, Q c , K c are multiplied after reshaping to obtain weight W2; The sorted features F b Reshaping is performed, and the reshaped features are multiplied by the weight W1 and the weight W2 respectively to obtain And Point-by-point multiplication is performed after reshaping to obtain F W ; F W Channel segmentation is performed to obtain features And The features are processed by scattering, and combined with indexes d1 and d2 to generate recovery features and features Splicing is performed, and the features F + are output through a 1*1 convolution layer. The first sorting is vertical sorting and horizontal sorting in sequence, and the second sorting is vertical sorting or horizontal sorting.

2. The remote sensing image segmentation method based on the intensity grouping Transformer network according to claim 1, characterized in that, The S1 comprises: S11, sequentially pre-processing the remote sensing image through a first preprocessing module, a second preprocessing module and a third preprocessing module to generate features f1, f2 and f3 respectively; and passing the remote sensing image through a 3*3 convolution layer and a first intensity grouping Transformer module group to generate a feature F1; S12, pixel recombination is performed on the feature F1, and feature splicing is performed with the feature f1 to generate a feature and the feature processed by a 1x1 convolutional layer and a second intensity grouping transformer module group to generate a feature F2; S13, pixel recombination is performed on the feature F2, and feature splicing is performed with the feature f2 to generate a feature and the feature The feature F3 is generated by processing through a 1x1 convolutional layer and a third intensity grouping Transformer module group. S14, pixel recombination is performed on the feature F3, and feature splicing is performed with the feature f3 to generate a feature and the feature The feature F4 is generated by processing through a 1x1 convolutional layer and a fourth intensity grouping Transformer module group.

3. The remote sensing image segmentation method based on the intensity group Transformer network according to claim 2, characterized in that, The first preprocessing module, the second preprocessing module and the third preprocessing module each comprise 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 the intensity grouping Transformer network according to claim 2, characterized in that, The pixel recombination comprises: dividing the feature into 2*2 non-overlapping blocks, flattening each block into 4 channel vectors and splicing to the channel dimension, increasing the channel number 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 the intensity group Transformer network according to claim 1, characterized in that, The S2 comprises: S21, inputting the feature F4 into the fifth intensity grouping Transformer module group, and performing pixel shuffling to generate a feature 4 ' F ; S22, feature 4 ' F Feature splicing is performed with feature F3, processed through a 1x1 convolutional layer and a sixth intensity grouping Transformer module group, and pixel shuffling is performed to generate feature 3 ' F ; S23, feature 3 ' F Feature splicing is performed with feature F2, processed by a 1x1 convolutional layer and a seventh intensity grouping Transformer module group, and pixel shuffling is performed to generate feature 2 ' F ; S24, the feature 2 ' F Feature splicing is performed with feature F1 to generate an initial segmentation result through a 1x1 convolutional layer, an eighth strength grouping Transformer module group, a 3x3 convolutional layer, and an activation function.

6. The remote sensing image segmentation method based on the intensity grouping Transformer network according to claim 5, characterized in that, The pixel shuffling includes: grouping low-resolution feature channel numbers C by r 2 times, and rearranging each group of channels into an r×r spatial block, so that the output feature spatial resolution is increased by r times and the channel number is reduced to C / r 2 , where r is a preset up-sampling rate.

7. The remote sensing image segmentation method based on the intensity group Transformer network according to claim 1, characterized in that, The S3 comprises: S31, using a conditional random field to optimize the boundary continuity of the initial segmentation result to enhance the edge definition of the ground object; S32, identifying a small target region through connected region analysis and performing a morphological dilation operation to enhance the integrity; S33, using a multi-scale Gaussian pyramid to perform feature fusion, extract detailed information at different scales and weight reconstruction to output a segmentation result corresponding to the remote sensing image.

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

Images