Remote sensing image segmentation method based on multi-layer wavelet transform and dynamic memory network

Abstract

The present invention relates to the field of image processing technology, and specifically to a remote sensing image segmentation method based on multi-layer wavelet transform and dynamic memory network, comprising: decomposing an image into a low-frequency sub-band and multiple high-frequency sub-bands through multi-layer Haar wavelet transform, and then extracting features from each frequency band in combination with convolution operation to obtain a final feature map; constructing a dynamic memory unit network, using a query-key-value structure to perform dynamic feature extraction on the feature map, generating dynamic memory enhancement features with global context information, and performing global-local feature modeling on the feature map, fusing the dynamic memory enhancement features with the global-local features to obtain a fused feature map; and segmenting the fused feature map to obtain a final segmentation result. The present invention can accurately extract low-frequency smoothing information and high-frequency detail information of the image, and can achieve efficient fusion of global and local features, thereby improving the accuracy and robustness of remote sensing image segmentation.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and more particularly to a remote sensing image segmentation method based on multi-layer wavelet transform and dynamic memory network. Background Art

[0002] With the rapid development of remote sensing technology, remote sensing images are playing an increasingly important role in fields such as land resources monitoring, urban planning, environmental protection, and disaster assessment. Remote sensing images, due to their wide acquisition range and large amount of information, often contain rich ground details and complex multi-scale features. Therefore, accurately extracting target information from this massive amount of data has become a key issue. Traditional image segmentation methods often perform poorly when dealing with complex backgrounds, low contrast, and multi-scale targets. While deep learning-based segmentation methods have improved this situation to some extent in recent years, they still suffer from limitations such as loss of local details, insufficient global information, large number of model parameters, and high computational complexity.

[0003] Therefore, how to improve the accuracy and robustness of remote sensing image segmentation is an urgent problem that needs to be solved by those skilled in the art. Summary of the Invention

[0004] In view of this, the present invention provides a remote sensing image segmentation method based on multi-layer wavelet transform and dynamic memory network, which can accurately extract low-frequency smooth information and high-frequency detail information of the image, and can realize the efficient fusion of global and local features, thereby improving the accuracy and robustness of remote sensing image segmentation.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] A remote sensing image segmentation method based on multi-layer wavelet transform and dynamic memory network includes the following steps:

[0007] S1. Based on the pre-trained Mamba model, the input remote sensing image is subjected to feature extraction. The image is decomposed into a low-frequency sub-band and multiple high-frequency sub-bands through multi-layer Haar wavelet transform. The convolution operation is then combined to extract features from each frequency band to obtain the final feature map φ. o ;

[0008] S2, build a dynamic memory unit network, use the query-key-value structure, and perform the feature map φ o Perform dynamic feature extraction to generate dynamic memory enhancement features with global context information and perform dynamic feature extraction on the feature map φ o Perform global-local feature modeling, fuse the dynamic memory enhancement features and global-local features to obtain the fused feature map φ + ;

[0009] S3, fusion feature map φ +Perform segmentation and obtain the final segmentation result.

[0010] Furthermore, S1 includes:

[0011] S11, extract features from the input remote sensing image based on the pre-trained Mamba model to obtain the initial feature map φ;

[0012] S12, perform a layer of Haar wavelet transform on the initial feature map φ, and then perform a convolution operation by combining a low-pass filter and a high-pass filter through a deep convolution operation to obtain a low-frequency sub-band and three high-frequency sub-bands;

[0013] S13, performing multi-layer wavelet transform on the low-frequency subband, and each layer of transform uses the same convolution operation to extract low-frequency features and high-frequency features;

[0014] S14, after multi-layer wavelet transform, using inverse wavelet transform to reconstruct each sub-band back to the low-frequency component and high-frequency component of the original image;

[0015] S15. Combine wavelet transform and convolution operation to extract features of the transformation results of each layer and obtain the final feature map φ o .

[0016] Furthermore, in S12, one low-frequency sub-band and three high-frequency sub-bands are represented as:

[0017] [φ L ,φ LH ,φ HL ,φ HH ]=Conv(f LL , f LH , f HL , f HH ],φ)

[0018] Among them, φ L represents the low-frequency subband, which contains the smoothing information of the image; φ LH Represents the horizontal high-frequency subband, which contains the horizontal edge information of the image; φ HL Represents the vertical high-frequency subband, which contains the vertical edge information of the image; φ HH Represents the diagonal high-frequency subband, which contains the diagonal edge information of the image; the resolution of each subband is half of the original image; f LL represents a low-pass filter, f LH represents the horizontal high-frequency filter, f HL represents the vertical high-frequency filter, f HH Represents a diagonal high-frequency filter; Conv represents a convolution operation.

[0019] Furthermore, in S13, the wavelet transform of each layer is expressed as:

[0020]

[0021] in, Represents the low-frequency subband of the kth layer, extracting the feature information of the image layer by layer until the predetermined number of layers is reached; DWT k Represents the wavelet transform operation of the kth layer. The decomposition of each layer generates new subbands, including LL k , LH k , HL k and HH k , which contain the information of the image at different frequencies.

[0022] Furthermore, S2 includes:

[0023] S21. Construct a key memory unit and a value memory unit through two independent and parameter-shared linear layers. The two memory units are used to extract and store deep feature information across regions and scales in remote sensing images.

[0024] S22, from the feature map φ o Extract the query weight Q, adopt the dynamic query mechanism, match the query weight Q with the stored features in the two memory units, and generate dynamic memory enhanced features with global context information;

[0025] S23, using adaptive average pooling and depth-separable convolution to feature map φ o Model the local information of the model to obtain local features;

[0026] S24. Calculate feature map φ o The learnable weight of the global feature is multiplied by the local feature to obtain the global-local feature;

[0027] S25. Perform weighted summation of global-local features and dynamic memory enhancement features to obtain a fusion feature map φ + .

[0028] Furthermore, in S22, the calculation formula of the query weight Q is:

[0029] Q=W q *Norm(Conv(φ o ))+b q

[0030] Among them, Conv(·) is responsible for local feature extraction; Norm represents the normalization operation; W q and b q To ensure that the model can automatically focus on key areas, the parameters can be learned;

[0031] The dynamic memory enhancement feature is represented as:

[0032]

[0033] Among them, φ MU represents dynamic memory enhancement features, softmax represents the normalized exponential function, Q represents the query weight, K represents the key, V represents the value, d k Indicates the dimension of the key.

[0034] Furthermore, in S23, the local features are expressed as:

[0035] φ DW =DWConv(AAP(φ o )

[0036] Among them, DWConv represents depth-wise separable convolution, AAP represents adaptive average pooling layer, φ DW Represents local features.

[0037] Furthermore, in S24, the learnable weights are expressed as:

[0038]

[0039] Among them, P represents the position embedding operation, which is used to supplement the geographic location information; FC represents the fully connected layer, and Sigmoid represents the Sigmoid activation function; As a learnable weight, it can dynamically adjust the contribution of global features;

[0040] The global-local feature is expressed as:

[0041]

[0042] in, Represents a matrix multiplication operation.

[0043] Furthermore, in S25, the fusion feature map φ + Expressed as:

[0044] φ + =λ1φ MU +λ2φ GL

[0045] Among them, λ1 and λ2 are learnable parameters that can adaptively adjust the proportion of global and local information.

[0046] Furthermore, S3 includes:

[0047] S31, fusion feature map φ + Perform dimensionality reduction and semantic feature extraction to obtain the reduced dimensionality feature map F;

[0048] S32, using a bilinear interpolation upsampling operation to restore the reduced-dimensional feature map F to the original spatial size, and then scoring the category of each pixel in the upsampled feature;

[0049] S33, performing a softmax operation on the category score of each pixel, calculating the probability of each pixel in each category, performing an argmax operation on the probability distribution of each pixel, determining its predicted category, and obtaining a segmentation result;

[0050] S34, using a post-processing method to refine the segmentation boundary of the segmentation result and smooth the edge of the target area.

[0051] It can be seen from the above technical solutions that compared with the prior art, the present invention has the following beneficial effects:

[0052] 1. When the present invention performs feature extraction on remote sensing images, the pre-trained Mamba model is first used to perform preliminary feature extraction on the remote sensing images to obtain a primary feature map with certain semantic information. Taking into account that remote sensing images contain both smooth low-frequency information (such as large areas of water, farmland) and rich high-frequency details (such as roads and building edges), a multi-layer Haar wavelet transform is used to decompose the image. By performing wavelet transform on the image, the image can be decomposed into low-frequency sub-bands and multiple high-frequency sub-bands, which can effectively capture the low-frequency smooth information and high-frequency detail features in the remote sensing image at different scales, thereby providing a richer and more accurate feature expression for subsequent image segmentation. In addition, further processing of each frequency band in combination with the convolution operation can greatly enhance the expressive power of the feature map, so that the detail information of complex objects can be fully retained, greatly improving the accuracy and robustness of the segmentation results, and providing a more accurate feature basis for subsequent segmentation tasks. The strategy of combining multi-layer wavelet transform with convolution in the present invention enables the model to take into account both overall and local information when facing large-scale changes and subtle textures, and realizes full utilization of multi-scale features.

[0053] 2. The present invention fully utilizes the synergistic advantages of multi-scale features and dynamic memory mechanisms. By constructing a dynamic memory unit network based on a query-key-value structure, it realizes the effective fusion of global and local features through an adaptive dynamic query mechanism. It can not only capture long-distance dependencies and cross-scale information in images, but also adaptively adjust the contribution of global background and local details, significantly enhancing the ability to distinguish complex landforms (such as urban buildings and rural roads). Compared with traditional methods, this innovative strategy makes up for the shortcomings of existing technologies in capturing multi-frequency, multi-scale information and complex structure modeling in remote sensing images, thereby significantly improving the segmentation results in terms of detail restoration and overall coherence. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] 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.

[0055] Figure 1 A flowchart of a remote sensing image segmentation method based on multi-layer wavelet transform and dynamic memory network provided by the present invention;

[0056] Figure 2 This is a flow chart of S1 of the present invention;

[0057] Figure 3 This is a flow chart of S2 of the present invention. DETAILED DESCRIPTION

[0058] 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.

[0059] like Figure 1 As shown, the embodiment of the present invention discloses a remote sensing image segmentation method based on multi-layer wavelet transform and dynamic memory network, comprising the following steps:

[0060] S1. Based on the pre-trained Mamba model, the input remote sensing image is subjected to feature extraction. The image is decomposed into a low-frequency sub-band and multiple high-frequency sub-bands through multi-layer Haar wavelet transform. The convolution operation is then combined to extract features from each frequency band to obtain the final feature map φ. o ;

[0061] S2, build a dynamic memory unit network, use the query-key-value structure, and perform the feature map φ o Perform dynamic feature extraction to generate dynamic memory enhancement features with global context information and perform dynamic feature extraction on the feature map φ o Perform global-local feature modeling, fuse the dynamic memory enhancement features and global-local features to obtain the fused feature map φ + ;

[0062] S3, fusion feature map φ + Perform segmentation and obtain the final segmentation result.

[0063] The above steps are further explained below.

[0064] S1. Based on the pre-trained Mamba model, the input remote sensing image is subjected to feature extraction. The image is decomposed into a low-frequency sub-band and multiple high-frequency sub-bands through multi-layer Haar wavelet transform. The convolution operation is then combined to extract features from each frequency band to obtain the final feature map φ. o ;like Figure 2 As shown, S1 specifically includes:

[0065] S11. Preliminary feature extraction: Extract features from the input remote sensing image based on the pre-trained Mamba model to obtain the initial feature map φ. Where H and W are the height and width of the image respectively, and C is the number of channels of the feature map.

[0066] S12, Haar wavelet transform: initial feature map A layer of Haar wavelet transform is performed, and then a deep convolution operation is performed in combination with a low-pass filter and a high-pass filter to obtain a low-frequency sub-band and three high-frequency sub-bands.

[0067] The operation of a wavelet transform layer is: in each dimension (e.g., width or height), a depthwise convolution is performed using the following convolution kernel:

[0068] Low-pass filter f LL :

[0069] Horizontal high frequency filter f LH :

[0070] Vertical high frequency filter f HL :

[0071] Diagonal high frequency filter f HH :

[0072] After performing the convolution operation, the image is decomposed into four sub-bands, including a low-frequency sub-band and three high-frequency sub-bands, through a downsampling operation with a step size of 2. It can be expressed as:

[0073] [φ L ,φ LH ,φ HL ,φ HH ]=Conv(f LL , f LH , f HL , f HH ],φ)

[0074] Among them, φ L represents the low-frequency subband, which contains the smoothing information of the image; φ LH Represents the horizontal high-frequency subband, which contains the horizontal edge information of the image; φHL Represents the vertical high-frequency subband, which contains the vertical edge information of the image; φ HH Represents the diagonal high-frequency subband, which contains the diagonal edge information of the image; the resolution of each subband is half of the original image; f LL represents a low-pass filter, f LH represents the horizontal high-frequency filter, f HL represents the vertical high-frequency filter, f HH Represents a diagonal high-frequency filter; Conv represents a convolution operation.

[0075] S13. Perform multi-layer wavelet transform on the low-frequency sub-band to obtain finer feature information. Each layer of transform uses the same convolution operation to extract low-frequency features and high-frequency features. The wavelet transform of each layer is expressed as:

[0076]

[0077] in, Represents the low-frequency subband of the kth layer, extracting the feature information of the image layer by layer until the predetermined number of layers is reached; DWT k Represents the wavelet transform operation of the kth layer. The decomposition of each layer generates new subbands, including LL k , LH k , HL k and HH k , which contain the information of the image at different frequencies.

[0078] S14. Reconstruct the image using inverse wavelet transform: After the multi-layer wavelet transform, the inverse wavelet transform (IWT) is used to reconstruct each sub-band back to the low-frequency and high-frequency components of the original image. The inverse wavelet transform can be achieved by transposed convolution:

[0079] φ=Trans([f LL ,f LH ,f HL ,f HH ],[φ L ,φ LH ,φ HL ,φ HH ])

[0080] This operation restores the four sub-bands to the form of the original image. Through inverse wavelet transform, an image with different frequency information can be obtained, which provides support for subsequent feature extraction.

[0081] S15. In order to improve the expressive power of the feature map and reduce the number of parameters, wavelet transform and convolution operation are combined to extract features of the transformation results of each layer, and convolution kernel is used to further extract features to obtain the final feature map φ o The formula is:

[0082] φ o =IWT(Conv(W,WT(φ)))

[0083] Where W is a k×k convolution kernel, and depthwise convolution is usually considered, with the number of input channels being four times that of the image. o is the final output feature map. This method separates different frequency components through convolution operations, increasing the receptive field while maintaining computational efficiency. The convolution operation in each frequency band helps further extract detailed information.

[0084] S2. Build a dynamic memory unit network with the following structure: Figure 3 As shown. Using the query-key-value structure, the feature graph φ o Perform dynamic feature extraction to generate dynamic memory enhancement features with global context information and perform dynamic feature extraction on the feature map φ o Perform global-local feature modeling, fuse the dynamic memory enhancement features and global-local features to obtain the fused feature map φ + S2 specifically includes:

[0085] S21. Memory Unit Construction: Key and value memories are constructed using two independent, parameter-shared linear layers. These two memory units are used to extract and store deep feature information across regions and scales in remote sensing images. To enhance the model's long-term memory capacity, DMUNet employs key and value memories to store and retrieve deep feature information from remote sensing images. These memory units are implemented using two independent linear layers. The parameters of these linear layers are learnable and shared across the entire dataset, ensuring model consistency across samples while reducing computational complexity. During model initialization, the parameters of the key and value memories are randomly generated, and their output dimension (i.e., memory unit size) is controlled by the hyperparameter memory_size, which is typically set to a small value (such as 64 or 128) to prevent overfitting and improve computational efficiency. During training, the parameters of the key and value memories are updated via backpropagation and an optimizer (such as Adam or SGD).

[0086] S22. Dynamic feature query: Using the query-key-value (QKV) structure, feature information is stored and dynamically extracted, enabling the network to efficiently utilize deep feature information across scales and regions.

[0087] From the feature map φ o Extract the query weight Q to match the global features stored in the memory unit. The calculation formula of the query weight Q is:

[0088] Q=W q *Norm(Conv(φ o))+b q

[0089] Among them, Conv(·) is responsible for local feature extraction; Norm represents the normalization operation; W q and b q The parameters are learnable, ensuring that the model can automatically focus on the key areas.

[0090] A dynamic query mechanism is used to match the query weight Q with the stored features in the two memory units to generate dynamic memory-enhanced features with global context information. The dynamic memory-enhanced features are expressed as:

[0091]

[0092] Among them, φ MU It represents the dynamic memory enhancement feature, which can effectively supplement the long-distance context information and improve the segmentation accuracy of the target area in the remote sensing image; softmax represents the normalized exponential function, Q represents the query weight, K represents the key, V represents the value, d k Indicates the dimension of the key, often used to scale the dot product attention mechanism to ensure numerical stability.

[0093] Then, combined with global-local feature modeling, the remote sensing image segmentation task not only relies on global information modeling (such as background consistency and object classification), but also needs to capture local details (such as edges and texture features). To this end, the present invention uses full-dimensional dynamic convolution to improve the local feature expression ability, as follows:

[0094] S23, local feature extraction: Local information is crucial for remote sensing targets (such as roads, buildings, water bodies, etc.). This paper uses adaptive average pooling and depth-wise separable convolution to extract local features from the feature map φ. o The local information of is modeled to obtain local features; the local features are expressed as:

[0095] φ DW =DWConv(AAP(φ o ))

[0096] Among them, DWConv represents depth-wise separable convolution, AAP represents adaptive average pooling layer, φ DW Representing local features. This convolutional structure can reduce computational complexity while retaining key details, making it easier for the model to distinguish small targets (such as ships, vehicles, etc.) in remote sensing images.

[0097] S24, global feature extraction: In order to ensure that the model has global context information, the present invention introduces dynamic adaptive weight calculation:

[0098]

[0099] Among them, P represents the position embedding operation, which is used to supplement the geographic location information; FC represents the fully connected layer, and Sigmoid represents the Sigmoid activation function; As a learnable weight, the contribution of global features can be dynamically adjusted.

[0100] After that, the learnable weights are multiplied by the local features to obtain the global-local features; the global-local features are expressed as:

[0101]

[0102] in, Represents a matrix multiplication operation.

[0103] S25. Memory enhancement and global-local feature fusion: In remote sensing image segmentation, objects of different scales (such as urban buildings vs. rural roads) require different degrees of global background information and local details. Therefore, the present invention performs a weighted fusion of memory enhancement features, global features, and local features to obtain a fused feature map φ + . Fusion feature map φ + Expressed as:

[0104] φ + =λ1φ MU +λ2φ GL

[0105] Among them, λ1 and λ2 are learnable parameters that can adaptively adjust the proportion of global and local information to ensure that the model maintains good generalization ability on different tasks and data sets. The final fusion feature φ + It has the following key advantages:

[0106] 1) Enhance global perception capabilities (dynamic memory units store cross-scale information and reduce category confusion).

[0107] 2) Enhance the expression of detailed features (local convolution structure improves edge detection capabilities).

[0108] 3) Improve robustness (multi-scale fusion strategy reduces the impact of illumination changes and noise interference on segmentation results).

[0109] After S2, the fusion feature φ is obtained + It fully integrates global memory and local detail information, and can accurately depict multi-scale objects in remote sensing images.

[0110] S3, fusion feature map φ + Segmentation is performed to obtain the final segmentation result. S3 is obtained from the fusion feature φ +The final pixel-level segmentation result is generated in . The whole process mainly includes feature map dimensionality reduction, upsampling prediction, normalization and necessary post-processing optimization. Specifically including:

[0111] S31, Feature Mapping Dimensionality Reduction and Semantic Fusion: In order to fusion feature φ + Convert it into an intermediate representation suitable for segmentation tasks, first perform dimensionality reduction and semantic information extraction. Specifically, a 1×1 convolutional layer is used for channel fusion, and a nonlinear activation function is used to further enhance the feature expression:

[0112] F=f(Conv 1×1 (φ + ))

[0113] Among them, Conv 1×1 represents a 1×1 convolution operation, used for dimensionality reduction and feature fusion; f(·) is the activation function (such as ReLU or GELU); and F represents the feature map after dimensionality reduction. This dimensionality reduction helps extract high-level semantic information while reducing computational complexity, laying the foundation for subsequent upsampling and segmentation prediction.

[0114] S32. Upsampling and segmentation prediction: Remote sensing images usually have high resolution, so the feature map F after dimensionality reduction must be restored to the original spatial size to ensure the accuracy of pixel-level segmentation results. Bilinear interpolation upsampling operation is used to expand F to a size that matches the input image, and bilinear upsampling operation is used to obtain Aligned with the height H and width W of the original image. Next, another 1×1 convolutional layer is used to further process the upsampled features to generate the category score for each pixel. represents the raw score of each pixel in all categories, where C cls is the number of categories.

[0115] S33, normalization and segmentation result generation: In order to convert the original score S of each pixel category into an easily interpretable probability distribution, a softmax operation is applied to S to calculate the probability of each pixel in each category. Represents the pixel-level category probability distribution. Finally, by performing an argmax operation on the probability distribution of each pixel, its predicted category is determined and the final segmentation result is obtained. This is the final pixel-level segmentation map, where each pixel (i, j) is assigned a specific category label.

[0116] S34. Use post-processing methods to refine the segmentation boundaries of the segmentation results and smooth the edges of the target area. For remote sensing image segmentation tasks, especially when dealing with complex object boundaries, the preliminary segmentation results may have noise or the edges are not smooth enough. To this end, the following post-processing techniques can be used to further optimize the segmentation results: Conditional Random Field (CRF), which refines the segmentation boundaries and improves the overall segmentation coherence by utilizing the color and position similarity between pixels. Morphological operations, such as erosion and dilation operations, help eliminate small noise and smooth the edges of the target area. These post-processing steps can be used selectively according to actual application requirements to obtain higher quality segmentation maps.

[0117] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

[0118] 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 multi-layer wavelet transform and dynamic memory network, characterized in that: The following steps are involved: S1. Based on the pre-trained Mamba model, the input remote sensing image is subjected to feature extraction. The image is decomposed into a low-frequency sub-band and multiple high-frequency sub-bands through multi-layer Haar wavelet transform. The convolution operation is then combined to extract features from each frequency band to obtain the final feature map φ. o ; S2, build a dynamic memory unit network, use the query-key-value structure, and perform the feature map φ o Perform dynamic feature extraction to generate dynamic memory enhancement features with global context information and perform dynamic feature extraction on the feature map φ o Perform global-local feature modeling, fuse the dynamic memory enhancement features and global-local features to obtain the fused feature map φ + ; S2 includes: S21. Construct a key memory unit and a value memory unit through two independent and parameter-shared linear layers. The two memory units are used to extract and store deep feature information across regions and scales in remote sensing images. S22, from the feature map φ o Extract the query weight Q, adopt the dynamic query mechanism, match the query weight Q with the stored features in the two memory units, and generate dynamic memory enhanced features with global context information; S23, using adaptive average pooling and depth-separable convolution to feature map φ o Model the local information of the model to obtain local features; S24. Calculate feature map φ o The learnable weight of the global feature is multiplied by the local feature to obtain the global-local feature; S25. Perform weighted summation of global-local features and dynamic memory enhancement features to obtain a fusion feature map φ + ; S3, fusion feature map φ + Perform segmentation and obtain the final segmentation result.

2. The remote sensing image segmentation method based on multi-layer wavelet transform and dynamic memory network according to claim 1, characterized in that: S1 includes: S11, extract features from the input remote sensing image based on the pre-trained Mamba model to obtain the initial feature map φ; S12, perform a layer of Haar wavelet transform on the initial feature map φ, and then perform a convolution operation by combining a low-pass filter and a high-pass filter through a deep convolution operation to obtain a low-frequency sub-band and three high-frequency sub-bands; S13, performing multi-layer wavelet transform on the low-frequency subband, and each layer of transform uses the same convolution operation to extract low-frequency features and high-frequency features; S14, after multi-layer wavelet transform, using inverse wavelet transform to reconstruct each sub-band back to the low-frequency component and high-frequency component of the original image; S15. Combine wavelet transform and convolution operation to extract features of the transformation results of each layer and obtain the final feature map φ o .

3. The remote sensing image segmentation method based on multi-layer wavelet transform and dynamic memory network according to claim 2, characterized in that: In S12, one low-frequency sub-band and three high-frequency sub-bands are represented as: [f L ,f LH ,f HL ,f HH ]=Conv([f LL ,f LH ,f HL ,f HH ],φ) Among them, φ L represents the low-frequency subband, which contains the smoothing information of the image; φ LH Represents the horizontal high-frequency subband, which contains the horizontal edge information of the image; φ HL Represents the vertical high-frequency subband, which contains the vertical edge information of the image; φ HH Represents the diagonal high-frequency subband, which contains the diagonal edge information of the image; the resolution of each subband is half of the original image; f LL represents a low-pass filter, f LH represents the horizontal high-frequency filter, f HL represents the vertical high-frequency filter, f HH Represents a diagonal high-frequency filter; Conv represents a convolution operation.

4. The remote sensing image segmentation method based on multi-layer wavelet transform and dynamic memory network according to claim 2, characterized in that: In S13, the wavelet transform of each layer is expressed as: in, Represents the low-frequency subband of the kth layer, extracting the feature information of the image layer by layer until the predetermined number of layers is reached; DWT k Represents the wavelet transform operation of the kth layer. The decomposition of each layer generates new subbands, including LL k ,LH k ,HL k and HH k , which contain the information of the image at different frequencies.

5. The remote sensing image segmentation method based on multi-layer wavelet transform and dynamic memory network according to claim 1, characterized in that: In S22, the calculation formula of the query weight Q is: Q=W q *Norm(Conv(φ o ))+b q Among them, Conv(·) is responsible for local feature extraction; Norm represents the normalization operation; W q and b q To ensure that the model can automatically focus on key areas, the parameters can be learned; The dynamic memory enhancement feature is represented as: Among them, φ MU represents dynamic memory enhancement features, softmax represents the normalized exponential function, Q represents the query weight, K represents the key, V represents the value, d k Indicates the dimension of the key.

6. The remote sensing image segmentation method based on multi-layer wavelet transform and dynamic memory network according to claim 1, characterized in that: In S23, the local features are expressed as: φ DW =DWConv(AAP(φ o )) Among them, DWConv represents depth-wise separable convolution, AAP represents adaptive average pooling layer, φ DW Represents local features.

7. The remote sensing image segmentation method based on multi-layer wavelet transform and dynamic memory network according to claim 6, characterized in that: In S24, the learnable weights are expressed as: Among them, P represents the position embedding operation, which is used to supplement the geographic location information; FC represents the fully connected layer, and Sigmoid represents the Sigmoid activation function; As a learnable weight, it can dynamically adjust the contribution of global features; The global-local feature is expressed as: in, Represents a matrix multiplication operation.

8. The remote sensing image segmentation method based on multi-layer wavelet transform and dynamic memory network according to claim 7, characterized in that: In S25, the fusion feature map φ + Expressed as: f + =λ1φ MU +λ2φ GL Among them, λ1 and λ2 are learnable parameters that can adaptively adjust the proportion of global and local information.

9. The remote sensing image segmentation method based on multi-layer wavelet transform and dynamic memory network according to claim 1, characterized in that S3 include: S31, fusion feature map φ + Perform dimensionality reduction and semantic feature extraction to obtain the reduced dimensionality feature map F; S32, using a bilinear interpolation upsampling operation to restore the reduced-dimensional feature map F to the original spatial size, and then scoring the category of each pixel in the upsampled feature; S33, performing a softmax operation on the category score of each pixel, calculating the probability of each pixel in each category, performing an argmax operation on the probability distribution of each pixel, determining its predicted category, and obtaining a segmentation result; S34, using a post-processing method to refine the segmentation boundary of the segmentation result and smooth the edge of the target area.

Images