Remote sensing image segmentation method, system and device and storage medium

Abstract

The invention relates to the technical field of remote sensing image processing and computer vision, in particular to a remote sensing image segmentation method, system and device and a storage medium, and the method comprises the steps: obtaining a to-be-segmented remote sensing image; inputting a remote sensing image to be segmented into the trained remote sensing image segmentation model to obtain a remote sensing image segmentation result; the trained remote sensing image segmentation model comprises an encoder, a decoder and a segmentation head; according to the encoder, parallel global branches and local branches are adopted to carry out feature extraction in stages, and encoding features are obtained; the decoder is used for carrying out feature decoding processing on the coding features in stages to obtain decoding features; performing feature fusion on the decoding feature of each stage and the encoding feature of the corresponding stage to obtain a fused feature; and the segmentation head maps the fusion features to obtain a segmentation result of the remote sensing image. The method has important significance in the fields of urban planning, land resource management, environment monitoring and the like.

Description

Technical Field

[0001] The present invention relates to the field of remote sensing image processing and computer vision technology, and in particular to a remote sensing image segmentation method, system, device and storage medium. Background Art

[0002] Semantic segmentation of remote sensing images aims to accurately classify each pixel in remote sensing images to automatically identify and extract land cover types, such as buildings, farmland, and water bodies. This technology is a key enabler for applications such as urban planning, resource management, and environmental monitoring. However, high-resolution remote sensing images contain rich ground feature information, diverse target types, and significant spatial scale differences. Furthermore, complex spatial relationships and contextual information must be accounted for, posing significant challenges for achieving accurate semantic segmentation. Traditional methods such as support vector machines (SVMs) and random forests require extensive manual intervention and have limited effectiveness. Deep learning methods based on convolutional neural networks (CNNs), such as U-Net and its variants, excel at extracting local features but are limited in modeling long-range image dependencies. Transformer-based methods, while using a self-attention mechanism, can effectively capture long-range contextual information. However, their computational complexity grows quadratically with the length of the input sequence, resulting in significant computational and memory overhead when processing high-resolution remote sensing images, hindering their practical application.

[0003] Recently, the Mamba model, a state-space model, has demonstrated its ability to efficiently model long sequences with linear computational complexity in natural language processing. Its visual variants, such as Vision Mamba, have begun to be introduced into computer vision tasks. However, existing Mamba-based visual models still have shortcomings in remote sensing image segmentation: Firstly, they inadequately extract specific global information in the spatial and channel dimensions; secondly, their ability to perceive local details needs improvement. Therefore, designing a lightweight semantic segmentation method that efficiently captures multi-dimensional global contextual information while maintaining sensitivity to local details is a key challenge that needs to be addressed. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a remote sensing image segmentation method, system, device, and storage medium. By designing a fusion spatial channel Mamba module and an information-guided cross-fusion module, this method effectively combines the state-space modeling capabilities of the Mamba model with its attention mechanism. While maintaining linear computational complexity, it significantly improves the ability to extract global contextual information (including spatial and channel dimensions) from remote sensing images and the perception of local details, thereby achieving pixel-level precision in the classification of land cover categories (such as buildings, roads, vegetation, and water bodies). This method has important implications for urban planning, land resource management, environmental monitoring, and other fields.

[0005] In one aspect, a remote sensing image segmentation method is provided, comprising: Obtaining a remote sensing image to be segmented; The remote sensing image to be segmented is input into a trained remote sensing image segmentation model to obtain a segmentation result of the remote sensing image; wherein the trained remote sensing image segmentation model includes: an encoder, a decoder and a segmentation head; the encoder uses parallel global branches and local branches to perform feature extraction in stages to obtain encoding features; the decoder performs feature decoding processing on the encoding features in stages to obtain decoding features; the decoding features of each stage are fused with the encoding features of the corresponding stage to obtain fused features; the segmentation head maps the fused features to obtain a segmentation result of the remote sensing image.

[0006] In another aspect, a remote sensing image segmentation system is provided, comprising: An acquisition module is configured to: acquire a remote sensing image to be segmented; The segmentation module is configured to: input the remote sensing image to be segmented into a trained remote sensing image segmentation model to obtain the segmentation result of the remote sensing image; wherein the trained remote sensing image segmentation model includes: an encoder, a decoder and a segmentation head; the encoder uses parallel global branches and local branches to perform feature extraction in stages to obtain encoding features; the decoder performs feature decoding processing on the encoding features in stages to obtain decoding features; the decoding features of each stage are fused with the encoding features of the corresponding stage to obtain fused features; the segmentation head maps the fused features to obtain the segmentation result of the remote sensing image.

[0007] In another aspect, an electronic device is provided, comprising: a memory for non-transitory storage of computer-readable instructions; and a processor for executing said computer-readable instructions, When the computer-readable instructions are executed by the processor, the method described in the first aspect is executed.

[0008] On the other hand, a storage medium is provided, which non-temporarily stores computer-readable instructions, wherein when the non-temporary computer-readable instructions are executed by a computer, the method described in the first aspect is executed.

[0009] The above technical solution has the following advantages or beneficial effects: This paper adopts the classic encoder-decoder structure used in semantic segmentation tasks and designs a spatial channel Mamba module. This module combines Mamba's state-space model with spatial and channel attention mechanisms, applying them simultaneously in the encoding and decoding stages. During the encoding stage, it efficiently captures global feature representations in both spatial and channel dimensions, enhancing the model's ability to focus on specific dimensions. Simultaneously, during the decoding stage, it fuses the encoded information with the decoded information from the previous layer, effectively bridging the semantic gap between the encoder's skip connection features and the decoder's upsampled features while preserving fine detail. This method significantly improves the semantic segmentation accuracy of remote sensing images while maintaining low computational complexity. It exhibits excellent robustness and generalization capabilities, and has broad application prospects in fields such as urban planning and resource surveys. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] The accompanying drawings, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0011] Figure 1 This is a flow chart of the method of embodiment 1. DETAILED DESCRIPTION

[0012] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present invention belongs.

[0013] Example 1 This embodiment provides a remote sensing image segmentation method; like Figure 1 As shown, a remote sensing image segmentation method includes: S101: Acquire a remote sensing image to be segmented; S102: Input the remote sensing image to be segmented into the trained remote sensing image segmentation model to obtain the segmentation result of the remote sensing image; wherein the trained remote sensing image segmentation model includes: an encoder, a decoder and a segmentation head; the encoder uses parallel global branches and local branches to perform feature extraction in stages to obtain encoding features; the decoder performs feature decoding processing on the encoding features in stages to obtain decoding features; the decoding features of each stage are fused with the encoding features of the corresponding stage to obtain fused features; the segmentation head maps the fused features to obtain the segmentation result of the remote sensing image.

[0014] Furthermore, the training process of the trained remote sensing image segmentation model includes: Constructing a training set, wherein the training set is a remote sensing image with known image segmentation results; The training set is input into the remote sensing image segmentation model to train the model. When the total loss function value of the model no longer decreases, the training is stopped to obtain the trained remote sensing image segmentation model.

[0015] Furthermore, the input end of the encoder is used to input the remote sensing image to be segmented, the output end of the encoder is connected to the input end of the decoder, the output end of the decoder is connected to the input end of the segmentation head, and the output end of the segmentation head outputs the segmentation result of the remote sensing image.

[0016] Furthermore, the global branch includes: a block layer Embedding, a first spatial channel Mamba module, a first adder, a second spatial channel Mamba module, a first convolutional layer, a second adder, a third spatial channel Mamba module, a second convolutional layer, a third adder, a fourth spatial channel Mamba module, a third convolutional layer and a fourth adder connected in sequence; wherein, the input end of the block layer Embedding is used to input the remote sensing image to be segmented; the output end of the fourth adder is the output end of the encoder, and the output end of the fourth adder is connected to the input end of the decoder.

[0017] Furthermore, the blocking layer Embedding and the first spatial channel Mamba module constitute the first stage of the global branch, the second spatial channel Mamba module and the first convolutional layer constitute the second stage of the global branch, the third spatial channel Mamba module and the second convolutional layer constitute the third stage of the global branch, and the fourth spatial channel Mamba module and the third convolutional layer constitute the fourth stage of the global branch.

[0018] The working process formula of each stage of the global branch except the first stage is expressed as follows:

[0019] in, Indicates the first The output of the stage, Indicates the local branch The output of the stage, Indicates the Space channel Mamba module, Indicates the global branch The intermediate results of the stage, Indicates the Convolutional layers, Indicates the first Output of the stage.

[0020] The first stage in the global branch is slightly different from the second to fourth stages. The first stage consists of the block layer Embedding and the first spatial channel Mamba module. The working process formula of the first stage is expressed as:

[0021] in, represents the remote sensing image to be segmented, represents the preliminary features extracted from remote sensing images, Represents the first spatial channel Mamba module, represents the output of the first stage in the global branch, Indicates the embedding layer.

[0022] The first stage, the embedding layer, divides the input remote sensing image into non-overlapping image blocks and generates preliminary features. Specifically, the image is divided into non-overlapping 4×4 image blocks and the feature dimension is mapped to C. This operation is achieved using a convolutional layer with a kernel size of 4×4 and a stride of 4. This process reduces the spatial resolution to one-quarter of the original size, resulting in the feature output of the first stage having a dimension of H / 4×W / 4×C.

[0023] The second to fourth stages are then processed sequentially through the spatial channel Mamba module, ultimately generating a hierarchical global feature representation. To ensure that the size of features at each stage in the global branch is consistent with that in the local branch, the spatial channel Mamba module in the second to fourth stages is followed by a convolutional layer (with a kernel size of 3×3 and a stride of 2). The final feature dimensions of the second to fourth stages are H / 8×W / 8, H / 16×W / 16, and H / 32×W / 32, respectively.

[0024] The beneficial effects of the above technical solution are: the global branch can efficiently capture the global long-range dependencies in the spatial and channel dimensions, and simultaneously capture the cross-channel dependencies and long-range spatial context, significantly improving the model's ability to model the global context information of remote sensing images while maintaining linear computational complexity, overcoming the quadratic computational complexity of the traditional Transformer ( )question.

[0025] Furthermore, the local branch includes: a first residual module, a second residual module, a third residual module and a fourth residual module connected in sequence; the input end of the first residual module is used to input the remote sensing image to be segmented, and the output end of the fourth residual module is connected to the input end of the fourth adder.

[0026] The working process of the residual module is formulated as follows:

[0027] in, Indicates the The output of the residual module, Indicates the residual modules, Indicates the The output of each residual module. Each residual module consists of a different number of residual blocks, where the first residual module consists of three residual blocks, the second residual module consists of four residual blocks, the third residual module consists of six residual blocks, and the fourth residual module consists of three residual blocks. The first residual block of each residual module is a residual block with downsampling, and the non-first residual blocks of each residual module are residual blocks without downsampling.

[0028] The working process of the residual block with downsampling is formulated as follows:

[0029] in, represents the input of the residual block with downsampling, represents the residual block output with downsampling, represents a convolution layer with a kernel size of m and a stride of n. express activation function, Represents the intermediate result of the residual block with downsampling.

[0030] The working process of the residual block without downsampling is formulated as follows:

[0031] in represents the input of the residual block without downsampling, represents the output of the residual block without downsampling, represents a convolution layer with a kernel size of m and a stride of n. express activation function, Represents the intermediate result of the residual block without downsampling.

[0032] Furthermore, the output end of the first residual module is connected to the input end of the first adder, the output end of the second residual module is connected to the input end of the second adder, the output end of the third residual module is connected to the input end of the third adder, and the output end of the fourth residual module is connected to the input end of the fourth adder.

[0033] It should be understood that the input remote sensing image is , where H, W, and C are the height, width, and number of channels of the image, respectively. The first residual module, the second residual module, the third residual module, and the fourth residual module generate feature maps with spatial resolutions of H / 4×W / 4, H / 8×W / 8, H / 16×W / 16, and H / 32×W / 32, respectively.

[0034] The beneficial effects of the above technical solution are: the local branch efficiently captures pixel-level detail features through the residual network, and cooperates with downsampling to form a multi-scale feature pyramid. The residual structure helps to retain more local feature information and provide multi-level and more detailed local feature information for the subsequent decoder.

[0035] Furthermore, the decoder includes: a fifth spatial channel Mamba module, a first upsampling layer, a sixth spatial channel Mamba module, a second upsampling layer, a seventh spatial channel Mamba module, a third upsampling layer, an eighth spatial channel Mamba module and a fourth upsampling layer connected in sequence; the input end of the fifth spatial channel Mamba module is the input end of the decoder, and the fourth upsampling layer is the output end of the decoder.

[0036] Furthermore, the fifth spatial channel Mamba module and the first upsampling layer constitute the fourth stage of the decoder; the sixth spatial channel Mamba module and the second upsampling layer constitute the third stage of the decoder; the seventh spatial channel Mamba module and the third upsampling layer constitute the second stage of the decoder; and the eighth spatial channel Mamba module and the fourth upsampling layer constitute the first stage of the decoder.

[0037] The decoder, The encoding features of the corresponding level will be Compared with the upsampling result of the decoding feature in the previous stage The two inputs are fed into the spatial channel Mamba module for feature interaction and enhancement. Specifically, before being fed into the spatial channel Mamba module, the two inputs are initially fused by adding the corresponding positions, and then fed into the spatial channel Mamba module. The above operation can be formulated as:

[0038] in, Indicates the encoder The output of the stage, Indicates the decoder The output of the stage, Indicates the decoder The space channel Mamba module of the stage, Represents an upsampling operation.

[0039] It should be noted that the spatial channel Mamba module in the fourth stage of the decoder only has one encoded feature as input, so there is no feature addition operation.

[0040] The beneficial effect of the above technical solution is that the decoder dynamically fuses the encoder jump connection features and upsampling features during the decoding process through efficient integration of multi-level features, which not only retains high-resolution spatial details but also corrects semantic deviations through global context information.

[0041] Furthermore, the input end of the eighth spatial channel Mamba module is connected to the output end of the first adder, the input end of the seventh spatial channel Mamba module is connected to the output end of the second adder, the input end of the sixth spatial channel Mamba module is connected to the output end of the third adder, and the input end of the fifth spatial channel Mamba module is connected to the output end of the fourth adder.

[0042] Furthermore, the embedding layer is used to divide the input remote sensing image into non-overlapping image blocks and generate preliminary features.

[0043] Furthermore, the internal structures of the first spatial channel Mamba module, the second spatial channel Mamba module, the third spatial channel Mamba module, the fourth spatial channel Mamba module, the fifth spatial channel Mamba module, the sixth spatial channel Mamba module, the seventh spatial channel Mamba module, and the eighth spatial channel Mamba module are consistent; the first spatial channel Mamba module includes: The first layer normalization module, the first visual state space module, the fifth adder, the second layer normalization module, the channel attention module, the sixth adder, the third layer normalization module, the second visual state space module, the seventh adder, the fourth layer normalization module, the spatial attention module and the eighth adder are connected in sequence.

[0044] Furthermore, the working processes of the first spatial channel Mamba module, the second spatial channel Mamba module, the third spatial channel Mamba module, the fourth spatial channel Mamba module, the fifth spatial channel Mamba module, the sixth spatial channel Mamba module, the seventh spatial channel Mamba module, and the eighth spatial channel Mamba module are consistent; the working process formula of the first spatial channel Mamba module is expressed as: ; ; ; ; in, Indicates the The input of the spatial channel Mamba module, Indicates the The output of the spatial channel Mamba module, represents the first layer normalization module, represents the second layer normalization module, represents the third layer normalization module, represents the fourth layer normalization module, represents the first visual state space module, represents the second visual state space module, represents the channel attention module, represents the spatial attention module, represents the first intermediate result, represents the second intermediate result, Indicates the third intermediate result.

[0045] Furthermore, the internal working processes of the first visual state space module and the second visual state space module are consistent. The working process of the first visual state space module is expressed as follows: The feature sequence normalized by the layer As the input of the first visual state space module, let the state vector Indicates time The hidden state at , its state transfer and update formula are as follows:

[0046] in, and are the network learnable weights, is the activation function, is the hidden state at the previous moment, Indicates the features at each time step (or sequence position).

[0047] It should be understood that the visual state space module is a new sequence modeling method based on the state space model (SSM), which aims to solve the quadratic computational complexity caused by the self-attention mechanism in the traditional Transformer architecture ( )question.

[0048] Furthermore, the channel attention module includes: a global average pooling layer, a fully connected layer, a first activation function layer and a first multiplier connected in sequence; the input end of the global average pooling layer is the input end of the channel attention module; the output end of the first multiplier is the output end of the channel attention module; wherein, the input end of the multiplier is also connected to the input end of the global average pooling layer.

[0049] Furthermore, the channel attention module includes: Let the input be F First, use the global average pooling operation to extract the representative elements of the features in each channel , and then the channel attention weight is obtained through the full connection layer and Sigmoid function W , and then use the weight vector W With the original feature map Multiply element by element to dynamically adjust the contribution of each channel; the above steps are formulated as: ; ; ; in, represents the global average pooling operation, represents a fully connected operation, represents the Sigmoid function, Represents features enhanced by channel attention.

[0050] Furthermore, the spatial attention module includes: a maximum pooling layer, a fourth convolutional layer, a second activation function layer and a second multiplier connected in sequence; the input end of the maximum pooling layer is the input end of the spatial attention module, and the input end of the spatial attention module is also connected to the input end of the fourth convolutional layer through the average pooling layer; the input end of the spatial attention module is also connected to the input end of the second multiplier, and the output end of the second multiplier is the output end of the spatial attention module.

[0051] Furthermore, the spatial attention module includes: Let the input be F , first perform average pooling and maximum pooling on the channel dimension to obtain the spatial representation of the feature map as well as , after splicing as well as Perform convolution operation and pass Function gets the normalized spatial attention matrix , using spatial attention With input features Multiply pixel by pixel to get the weighted features ; The above steps are formulated as: ; ; ; ; in, represents the channel average pooling operation, represents the channel maximum pooling operation, represents the convolution operation, represents the Sigmoid function, Represents features enhanced by spatial attention.

[0052] Furthermore, the first-layer normalization module, the second-layer normalization module, the third-layer normalization module, and the fourth-layer normalization module are used to accelerate training convergence and improve model stability and generalization ability.

[0053] Furthermore, the first residual module includes: a first residual unit with downsampling, a first residual unit without downsampling, and a second residual unit without downsampling, which are connected in sequence.

[0054] Furthermore, the second residual module includes: a second residual unit with downsampling, a third residual unit without downsampling, a fourth residual unit without downsampling, and a fifth residual unit without downsampling.

[0055] Furthermore, the third residual module includes: a third residual unit with downsampling, a sixth residual unit with downsampling, a seventh residual unit without downsampling, an eighth residual unit without downsampling, a ninth residual unit without downsampling, and a tenth residual unit without downsampling.

[0056] Furthermore, the fourth residual module includes: a fourth residual unit with downsampling, an eleventh residual unit with downsampling, and a twelfth residual unit without downsampling.

[0057] Furthermore, the first residual unit without downsampling, the second residual unit without downsampling, the third residual unit without downsampling, the fourth residual unit without downsampling, the fifth residual unit without downsampling, the sixth residual unit without downsampling, the seventh residual unit without downsampling, the eighth residual unit without downsampling, the ninth residual unit without downsampling, the tenth residual unit without downsampling, the eleventh residual unit without downsampling, and the twelfth residual unit without downsampling have the same internal structure, and the first residual unit without downsampling includes: A fifth convolutional layer, a sixth convolutional layer, and a ninth adder are connected in sequence; the input end of the fifth convolutional layer is the input end of the first down-sampled residual unit, the output end of the ninth adder is the output end of the first down-sampled residual unit, and the input end of the ninth adder is also connected to the input end of the fifth convolutional layer; Furthermore, the internal structures of the first down-sampled residual unit, the second down-sampled residual unit, the third down-sampled residual unit, and the fourth down-sampled residual unit are consistent, and the first down-sampled residual unit includes: The seventh convolutional layer, the eighth convolutional layer, and the tenth adder are connected in sequence. The input of the seventh convolutional layer is the input of the first residual unit without downsampling, and the output of the tenth adder is the output of the first residual unit without downsampling. The input of the seventh convolutional layer is also connected to the input of the tenth adder through the ninth convolutional layer. Downsampling is implemented by the ninth convolutional layer.

[0058] The beneficial effect of the above technical solution is: through the residual structure, the residual network can efficiently capture pixel-level detail features, and cooperate with downsampling to form a multi-scale feature pyramid. The residual structure helps to retain more local feature information and provide multi-level and more detailed local feature information for the subsequent decoder.

[0059] In summary, at each encoder layer, a global branch containing a spatial channel Mamba module and a local branch containing a residual module are used to extract multi-level features in parallel. At each decoder layer, the encoded features of the corresponding layer are combined with the upsampled decoded features of the previous layer and fed into the spatial channel Mamba module for feature interaction and enhancement. The fused features are then upsampled to generate multi-scale features with restored resolution. The features output by the final decoder layer are upsampled and mapped to the category space through the segmentation head (convolutional layer) to generate pixel-level semantic segmentation results.

[0060] In the decoder, the encoding features of the corresponding layer are input into the spatial channel Mamba module together with the upsampling results of the decoding features of the previous layer for feature interaction and enhancement. The fused features are upsampled to generate multi-scale features with restored resolution.

[0061] Decoder The encoding features of the corresponding level will be Compared with the upsampling result of the decoding feature in the previous stage The two inputs are fed into the spatial channel Mamba module for feature interaction and enhancement. Specifically, before being fed into the spatial channel Mamba module, the two inputs are initially fused by adding the corresponding positions, and then fed into the spatial channel Mamba module. The above operation can be formulated as:

[0062] in, Indicates the encoder i Stage output, Indicates the decoder i Stage output, Indicates the decoderi The space channel Mamba module of the stage, Represents an upsampling operation.

[0063] It should be noted that the spatial channel Mamba module in the fourth stage of the decoder only has one encoded feature as input, so there is no feature addition operation.

[0064] By encoding and decoding features, this method can effectively capture the global contextual information in remote sensing images. Whether it is small area details in the image or large-scale global features, the present invention can effectively model and learn them, providing more robust feature support for the subsequent generation of semantic segmentation results.

[0065] After decoding and upsampling at each stage of the decoder, the output features of the first stage of the decoder are The space of the image is restored to 1 / 4 (H / 4×W / 4). In order to obtain a segmentation result with the same resolution (H×W) as the input image, final upsampling and classification are required.

[0066] Furthermore, the segmentation head upsamples the final features output by the decoding layer and maps them to the category space through the segmentation head (convolutional layer) to generate pixel-level semantic segmentation results.

[0067] The features The spatial resolution is upsampled by 4 times through an upsampling layer to restore the original input image size H×W to obtain the feature map , where D is the number of feature channels. The upsampling layer is implemented by bilinear interpolation; Get feature map Then, the classification head (the classification head is implemented by the convolution layer) is used to map it to the number of channels of semantic categories to generate the semantic segmentation result; let the number of semantic categories be , the classification head performs 1×1 convolution on the final features Perform channel mapping to obtain features , ; in, Represents 1×1 convolution; The mapped features Apply the Softmax function for normalization to obtain the predicted probability of each category: ; Among the features The dimension is , and Represents the feature map middle The colon indicates the position All values ​​of the channels, The function normalizes the classification result of each pixel into a probability distribution, ensuring that the sum is 1. Represents pixels The probability of belonging to each category; For each pixel, the category with the highest probability is taken as its predicted semantic label, and finally a pixel-level semantic segmentation result map of the entire image is generated. , the above algorithm can be formulated as:

[0068] in, The function represents the category corresponding to the maximum probability. Represents the predicted semantic category of each pixel.

[0069] The predicted semantic categories of all pixels Combined into a complete semantic segmentation result map according to position , where each pixel value represents the object category label to which it belongs.

[0070] This invention innovatively combines the Mamba state-space model with spatial and channel attention mechanisms by designing a spatial channel Mamba module. This effectively captures global long-range dependencies in the spatial and channel dimensions during the encoding phase, significantly improving the model's ability to model the global contextual information of remote sensing images while maintaining linear computational complexity and overcoming the quadratic complexity bottleneck of the traditional Transformer. During the decoding phase, the spatial channel Mamba module is utilized to effectively fuse features from different sources (encoder jump features, decoder upsampling features) and of different properties (local details, global context). While retaining the fine structural information of the image, it significantly bridges the semantic gap between features and improves the recognition accuracy of complex scenes and small targets. This method significantly improves the semantic segmentation accuracy of remote sensing images while maintaining high computational efficiency, providing reliable technical support for applications such as land cover classification, urban planning, and disaster assessment.

[0071] Example 2 This embodiment provides a remote sensing image segmentation system, including: An acquisition module is configured to: acquire a remote sensing image to be segmented; The segmentation module is configured to: input the remote sensing image to be segmented into a trained remote sensing image segmentation model to obtain the segmentation result of the remote sensing image; wherein the trained remote sensing image segmentation model includes: an encoder, a decoder and a segmentation head; the encoder uses parallel global branches and local branches to perform feature extraction in stages to obtain encoding features; the decoder performs feature decoding processing on the encoding features in stages to obtain decoding features; the decoding features of each stage are fused with the encoding features of the corresponding stage to obtain fused features; the segmentation head maps the fused features to obtain the segmentation result of the remote sensing image.

[0072] It should be noted that the acquisition module and segmentation module described above correspond to steps S101 to S102 in Example 1. The examples and application scenarios implemented by the modules and the corresponding steps are the same, but are not limited to the contents disclosed in Example 1. It should be noted that the modules described above, as part of a system, can be executed in a computer system, such as a set of computer-executable instructions.

[0073] The description of each embodiment in the above embodiments has different emphases. For parts not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0074] The proposed system can be implemented in other ways. For example, the system embodiment described above is merely illustrative. For example, the above module division is only a logical function division. In actual implementation, other division methods may be used. For example, multiple modules can be combined or integrated into another system, or some features can be ignored or not implemented.

[0075] Example 3 This embodiment also provides an electronic device, comprising: one or more processors, one or more memories, and one or more computer programs; wherein the processor is connected to the memory, and the one or more computer programs are stored in the memory. When the electronic device is running, the processor executes the one or more computer programs stored in the memory, so that the electronic device executes the method described in the above embodiment one.

[0076] It should be understood that in this embodiment, the processor may be a central processing unit (CPU), or may be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), off-the-shelf field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc.

[0077] The memory may include a read-only memory and a random access memory, and provides instructions and data to the processor. A portion of the memory may also include a non-volatile random access memory. For example, the memory may also store information about the device type.

[0078] During implementation, each step of the above method may be completed by an integrated logic circuit of hardware in a processor or by instructions in the form of software.

[0079] The method in Example 1 can be directly implemented as being executed by a hardware processor, or by a combination of hardware and software modules within the processor. The software module can be located in a storage medium well-established in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. The storage medium is located in the memory, and the processor reads the information in the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, a detailed description is not given here.

[0080] Those skilled in the art will appreciate that the units and algorithm steps of the various examples described in conjunction with this embodiment can be implemented using electronic hardware or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present invention.

[0081] Example 4 This embodiment further provides a computer-readable storage medium for storing computer instructions. When the computer instructions are executed by a processor, the method described in the first embodiment is performed.

[0082] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

Claims

1. A remote sensing image segmentation method, characterized in that: include: Obtaining a remote sensing image to be segmented; The remote sensing image to be segmented is input into a trained remote sensing image segmentation model to obtain a segmentation result of the remote sensing image; wherein the trained remote sensing image segmentation model includes: an encoder, a decoder and a segmentation head; the encoder uses parallel global branches and local branches to perform feature extraction in stages to obtain encoding features; the decoder performs feature decoding processing on the encoding features in stages to obtain decoding features; the decoding features of each stage are fused with the encoding features of the corresponding stage to obtain fused features; the segmentation head maps the fused features to obtain a segmentation result of the remote sensing image.

2. A remote sensing image segmentation method according to claim 1, characterized in that: The global branch includes: a block layer Embedding, a first spatial channel Mamba module, a first adder, a second spatial channel Mamba module, a first convolutional layer, a second adder, a third spatial channel Mamba module, a second convolutional layer, a third adder, a fourth spatial channel Mamba module, a third convolutional layer, and a fourth adder connected in sequence; wherein the input end of the block layer Embedding is used to input the remote sensing image to be segmented; the output end of the fourth adder is the output end of the encoder, and the output end of the fourth adder is connected to the input end of the decoder; The blocking layer Embedding and the first spatial channel Mamba module constitute the first stage of the global branch, the second spatial channel Mamba module and the first convolutional layer constitute the second stage of the global branch, the third spatial channel Mamba module and the second convolutional layer constitute the third stage of the global branch, and the fourth spatial channel Mamba module and the third convolutional layer constitute the fourth stage of the global branch.

3. A remote sensing image segmentation method as claimed in claim 2, characterized in that: The local branch includes: a first residual module, a second residual module, a third residual module and a fourth residual module connected in sequence; the input end of the first residual module is used to input the remote sensing image to be segmented, and the output end of the fourth residual module is connected to the input end of the fourth adder; The output end of the first residual module is connected to the input end of the first adder, the output end of the second residual module is connected to the input end of the second adder, the output end of the third residual module is connected to the input end of the third adder, and the output end of the fourth residual module is connected to the input end of the fourth adder.

4. A remote sensing image segmentation method according to claim 3, characterized in that: The decoder includes: a fifth spatial channel Mamba module, a first upsampling layer, a sixth spatial channel Mamba module, a second upsampling layer, a seventh spatial channel Mamba module, a third upsampling layer, an eighth spatial channel Mamba module, and a fourth upsampling layer connected in sequence; an input end of the fifth spatial channel Mamba module is an input end of the decoder, and the fourth upsampling layer is an output end of the decoder; The fifth spatial channel Mamba module and the first upsampling layer constitute the fourth stage of the decoder; the sixth spatial channel Mamba module and the second upsampling layer constitute the third stage of the decoder; the seventh spatial channel Mamba module and the third upsampling layer constitute the second stage of the decoder; the eighth spatial channel Mamba module and the fourth upsampling layer constitute the first stage of the decoder; The input end of the eighth spatial channel Mamba module is connected to the output end of the first adder, the input end of the seventh spatial channel Mamba module is connected to the output end of the second adder, the input end of the sixth spatial channel Mamba module is connected to the output end of the third adder, and the input end of the fifth spatial channel Mamba module is connected to the output end of the fourth adder.

5. A remote sensing image segmentation method according to claim 2, characterized in that: The first spatial channel Mamba module includes: A first layer normalization module, a first visual state space module, a fifth adder, a second layer normalization module, a channel attention module, a sixth adder, a third layer normalization module, a second visual state space module, a seventh adder, a fourth layer normalization module, a spatial attention module, and an eighth adder connected in sequence; The working process formula of the first spatial channel Mamba module is expressed as: ； ； ； ； in, Indicates the The input of the spatial channel Mamba module, Indicates the The output of the spatial channel Mamba module, represents the first layer normalization module, represents the second layer normalization module, represents the third layer normalization module, represents the fourth layer normalization module, represents the first visual state space module, represents the second visual state space module, represents the channel attention module, represents the spatial attention module, represents the first intermediate result, represents the second intermediate result, Indicates the third intermediate result.

6. A remote sensing image segmentation method according to claim 5, characterized in that: The working process of the first visual state space module is expressed as follows: The feature sequence normalized by the layer As the input of the first visual state space module, let the state vector Indicates time t The hidden state at , its state transfer and update formula are as follows: ； in, and are the network learnable weights, is the activation function, is the hidden state at the previous moment, Represents the features of the t-th time step.

7. A remote sensing image segmentation method according to claim 5, characterized in that: The channel attention module includes: a global average pooling layer, a fully connected layer, a first activation function layer, and a first multiplier connected in sequence; the input end of the global average pooling layer is the input end of the channel attention module; the output end of the first multiplier is the output end of the channel attention module; wherein the input end of the multiplier is also connected to the input end of the global average pooling layer; The spatial attention module includes: a maximum pooling layer, a fourth convolutional layer, a second activation function layer and a second multiplier connected in sequence; the input end of the maximum pooling layer is the input end of the spatial attention module, and the input end of the spatial attention module is also connected to the input end of the fourth convolutional layer through the average pooling layer; the input end of the spatial attention module is also connected to the input end of the second multiplier, and the output end of the second multiplier is the output end of the spatial attention module.

8. A remote sensing image segmentation system, characterized in that: include: An acquisition module is configured to: acquire a remote sensing image to be segmented; The segmentation module is configured to: input the remote sensing image to be segmented into a trained remote sensing image segmentation model to obtain the segmentation result of the remote sensing image; wherein the trained remote sensing image segmentation model includes: an encoder, a decoder and a segmentation head; the encoder uses parallel global branches and local branches to perform feature extraction in stages to obtain encoding features; the decoder performs feature decoding processing on the encoding features in stages to obtain decoding features; the decoding features of each stage are fused with the encoding features of the corresponding stage to obtain fused features; the segmentation head maps the fused features to obtain the segmentation result of the remote sensing image.

9. An electronic device, comprising: a memory for non-transitory storage of computer-readable instructions; as well as a processor for executing said computer-readable instructions, When the computer-readable instructions are executed by the processor, the method according to any one of claims 1 to 7 is executed.

10. A storage medium, characterized in that: Non-transitory storage of computer-readable instructions, wherein when the non-transitory computer-readable instructions are executed by a computer, the method according to any one of claims 1 to 7 is performed.

Images