Selective state space model road segmentation method based on frequency domain feature compensation

Abstract

The invention discloses a selective state space model road segmentation method based on frequency domain feature compensation. The method comprises the following steps: acquiring a high-resolution road remote sensing image; the high-resolution road remote sensing image is sequentially input into four encoders with different resolutions, four different-resolution characteristics corresponding to the high-resolution road remote sensing image are obtained, and the four different-resolution characteristics include the first resolution characteristic, the second resolution characteristic, the third resolution characteristic and the fourth resolution characteristic; and inputting the four different resolution features into four corresponding spatial perception Mangbar modules to obtain four different spatial features, which are respectively a first spatial feature, a second spatial feature, a third spatial feature and a fourth spatial feature. According to the invention, the technical problem of inaccurate road segmentation caused by similarity and shielding problems of roads and backgrounds in high-resolution road remote sensing images in the prior art is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a selective state space model road segmentation method based on frequency domain feature compensation. Background Art

[0002] High-precision road maps are essential for multiple application scenarios, such as autonomous driving, urban planning, vehicle navigation, and geographic information systems. Traditional road map creation methods, such as LiDAR point cloud analysis, GPS trajectory integration, and manual annotation, although successful to some extent, still face challenges when dealing with large-scale or rapidly changing road networks; in contrast, satellite imagery, as an emerging road extraction method, can capture the geometric structure of roads and provide multi-temporal or real-time data, making it an ideal data source for efficient network extraction and updating; however, segmenting roads from high-resolution remote sensing images still faces many challenges, such as the non-uniformity of road texture, the fuzziness of pixels, the similarity with natural features such as rivers, and the presence of small-scale road objects in complex backgrounds; despite these difficulties, the importance of road segmentation in mapping, traffic management, and intelligent transportation has driven the research on continuous improvement of algorithms.

[0003] In recent years, many deep learning-based road segmentation methods have been proposed; compared with traditional methods, convolutional neural networks (CNNs) such as SegNet, U-Net and their variants have achieved promising results in road extraction. More recent models, such as D-LinkNet and CondiNet++, have adopted advanced convolution techniques such as conditional dilated convolutions to enhance the ability to extract fine-grained road features from high-resolution images; however, CNNs still have limitations in capturing long-range continuous structures and weak edges of roads in remote sensing images; models using strip convolutions and attention mechanisms, although they have alleviated this problem to some extent, still have difficulties in detecting roads that are occluded, small in size, or similar to the surrounding environment.

[0004] To overcome these difficulties, Transformer-based models were introduced to better model global and local contextual information. Although such models can capture a wide range of contextual information, the self-attention mechanism often acts as a low-pass filter, weakening high-frequency details, thereby blurring the detected road details. In addition, when processing high-resolution images, such models have high computational costs, which severely limits their feasibility and efficiency in actual road segmentation applications.

[0005] To solve these problems, the Mamba architecture has gradually attracted attention in the field of remote sensing. This architecture has efficient global information modeling capabilities and linear complexity, making it a powerful tool for further improving road extraction models; Mamba's structured state space model, combined with its selective scanning mechanism, can effectively establish correlations between image blocks, which is crucial in road segmentation because roads often present continuous structures over a large area; Mamba's modeling ability for long sequence relationships is particularly suitable for extracting road networks from high-resolution remote sensing images, where roads usually span wide areas and have complex patterns.

[0006] Although the Mamba architecture excels in capturing large-scale spatial patterns, it still has shortcomings in handling fine-grained details, especially the abrupt changes in road edges, occlusions, and the distinction between visually similar elements such as rivers or shadows. These subtle but critical details are difficult to capture through pure spatial domain analysis. Such details are often more obvious in frequency domain analysis. Summary of the invention

[0007] The embodiment of the present invention provides a selective state space model road segmentation method based on frequency domain feature compensation, so as to at least solve the technical problem of inaccurate road segmentation caused by the similarity and occlusion of roads and backgrounds in high-resolution road remote sensing images in the prior art.

[0008] According to one aspect of an embodiment of the present invention, a selective state space model road segmentation method based on frequency domain feature compensation is provided. The method may include: obtaining a high-resolution road remote sensing image; inputting the high-resolution road remote sensing image into four encoders with different resolutions in sequence to obtain four different resolution features corresponding to the high-resolution road remote sensing image, wherein the four different resolution features are respectively a first resolution feature, a second resolution feature, a third resolution feature and a fourth resolution feature; inputting the four different resolution features into the corresponding four spatial perception mamba modules to obtain four different spatial features, wherein the four different spatial features are respectively a first spatial feature, a second spatial feature, a third spatial feature and a fourth spatial feature; inputting the four different spatial features into the corresponding four wavelet transform modules to obtain four pairs of different target high-frequency components and low-frequency components, wherein the four pairs of different target high-frequency components and low-frequency components are respectively a first target high-frequency component and a first low-frequency component, a second target high-frequency component and a second low-frequency component, a third target high-frequency component and a third low-frequency component, and a fourth target high-frequency component and a fourth low-frequency component; the target high-frequency component of the fourth spatial feature The fourth high-frequency feature and the fourth low-frequency feature are output as the output of the fourth frequency domain feature compensation module, and the fourth high-frequency feature and the fourth low-frequency feature are output; the fourth high-frequency feature and the fourth low-frequency feature, the third target high-frequency component and the third low-frequency component are input into the third frequency domain feature compensation module to obtain the third high-frequency feature and the third low-frequency feature, and the process is repeated forward to obtain the second high-frequency feature and the second low-frequency feature of the second frequency domain feature compensation module, and the first high-frequency feature and the first low-frequency feature of the first frequency domain feature compensation module respectively; the fourth resolution feature is input into the feature after the void space convolution pooling pyramid, the fourth spatial feature and the first low-frequency feature are input into the fourth decoder to obtain the fourth decoded feature, the fourth decoded feature and the third spatial feature are input into the third decoder to obtain the third decoded feature, and the process is repeated forward to obtain the second decoded feature; the first high-frequency feature, the first spatial feature and the second decoded feature are input into the first decoder to obtain the first decoded feature; the first decoded feature is sequentially input into the convolution layer and the sigmoid function to obtain the segmentation result of the high-resolution road remote sensing image.

[0009] Optionally, the step of inputting four features of different resolutions into corresponding four spatial perception mamba modules to obtain four different spatial features includes: inputting four features of different resolutions into a pyramid pooling layer to obtain pooling features of four different sizes; upsampling the pooling features of four different sizes to the same size as the four features of different resolutions; splicing the upsampled features of the pooling features of four different sizes with the four features of different resolutions to obtain spliced ​​features of the four features of different resolutions; and inputting the spliced ​​features of the four features of different resolutions into a visual state space module to obtain four different spatial features.

[0010] Optionally, the four different spatial features are input into corresponding four wavelet transform modules to obtain four pairs of different target high-frequency components and low-frequency components, including: passing the four different spatial features through four wavelet transform modules to obtain four pairs of three different high-frequency components and one low-frequency component; connecting the four pairs of three different high-frequency components to obtain four pairs of different target high-frequency components.

[0011] Optionally, the fourth high-frequency feature and the fourth low-frequency feature, the third target high-frequency component and the third low-frequency component are input into the third frequency domain feature compensation module to obtain the third high-frequency feature and the third low-frequency feature, including: filtering the fourth high-frequency feature through a high-pass filter in the third frequency domain feature compensation module to obtain a fourth high-frequency feature map; filtering the fourth low-frequency feature through a low-pass filter in the third frequency domain feature compensation module to obtain a fourth low-frequency feature map; processing the fourth high-frequency feature map, the third target high-frequency component and the fourth high-frequency feature through a window-based linear model in the third frequency domain feature compensation module, and inputting them into a linear combination layer with the third target high-frequency component to obtain a third high-frequency feature; processing the fourth low-frequency feature map, the third low-frequency component and the fourth low-frequency feature through a window-based linear model in the third frequency domain feature compensation module, and inputting them into a linear combination layer with the third low-frequency component to obtain a third low-frequency feature.

[0012] Optionally, the backbone networks of the four encoders with different resolutions are all Mamba.

[0013] Optionally, each decoder consists of a strip convolution and upsampling module.

[0014] Beneficial effects of the present invention: The present invention proposes a selective state space model road segmentation method based on frequency domain feature compensation. Firstly, the proposed progressive frequency feature compensation method strengthens the prediction result of the road segmentation network, improves the segmentation performance, and promotes the development of related applications. Secondly, the frequency feature compensation module (FFC) decomposes the image features through wavelet transform to obtain high-frequency and low-frequency features, and supplements the high-frequency and low-frequency information in different network layers through progressive fusion. The high-frequency features are introduced into the shallow network to enhance the expression of details, while the low-frequency features are integrated into the deep network to enhance the representation ability of the overall road features. Then, the spatial perception Mamba module (SPMB) is proposed to enhance the expression ability of road spatial features without destroying spatial information. Finally, by combining the progressive frequency feature compensation method with the SPMB module, the present invention effectively solves the inaccurate segmentation phenomenon caused by the similarity and occlusion problems between the road and the background, and achieves a balance between global information capture and computational complexity when processing high-resolution remote sensing images. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings: Figure 1 is a flow chart of a selective state space model road segmentation method based on frequency domain feature compensation according to an embodiment of the present invention; Figure 2 4 is a block diagram of a selective state space model road segmentation method based on frequency domain feature compensation according to an embodiment of the present invention. DETAILED DESCRIPTION

[0016] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only embodiments of a part of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0017] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects and to describe a specific order or sequence. It should be understood that the terms used in this way can be interchangeable where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0018] Example 1 According to an embodiment of the present invention, a selective state-space model road segmentation method based on frequency domain feature compensation is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system comprising at least one set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in an order different from that shown here.

[0019] Figure 1 is a flow chart of a selective state space model road segmentation method based on frequency domain feature compensation according to an embodiment of the present invention. Figure 1 As shown, the method may include the following steps: Step S101, obtaining a high-resolution road remote sensing image.

[0020] In the technical solution provided in the above step S101 of the present invention, a high-resolution road remote sensing image is obtained, and the remote sensing image is an image including the background and the road.

[0021] Step S102, inputting the high-resolution road remote sensing image into four encoders with different resolutions in sequence to obtain four different resolution features corresponding to the high-resolution road remote sensing image, wherein the four different resolution features are a first resolution feature, a second resolution feature, a third resolution feature and a fourth resolution feature.

[0022] In the technical solution provided in the above step S102 of the present invention, Figure 2 is a block diagram of a selective state space model road segmentation method based on frequency domain feature compensation according to an embodiment of the present invention, such as Figure 2 As shown in the figure, four encoders with different resolutions are encoder 1, encoder 2, encoder 3 and encoder 4, which convert high-resolution road remote sensing images Input into the backbone network, and extract four features of different resolutions through the backbone network ,in , the backbone network includes encoder 1, encoder 2, encoder 3 and encoder 4.

[0023] Step S103, inputting four different resolution features into corresponding four spatial perception Mamba modules to obtain four different spatial features, wherein the four different spatial features are respectively a first spatial feature, a second spatial feature, a third spatial feature and a fourth spatial feature.

[0024] In the technical solution provided in step S103 of the present invention, Figure 2 As shown, the first resolution feature is input into the first spatial perception Mamba module of the first stage to obtain the first spatial feature, the second resolution feature is input into the second spatial perception Mamba module of the second stage to obtain the second spatial feature, the third resolution feature is input into the third spatial perception Mamba module of the third stage to obtain the third spatial feature, and the fourth resolution feature is input into the fourth spatial perception Mamba module of the fourth stage to obtain the fourth spatial feature.

[0025] Step S104, inputting four different spatial features into corresponding four wavelet transform modules to obtain four pairs of different target high-frequency components and low-frequency components, wherein the four pairs of different target high-frequency components and low-frequency components are respectively a first target high-frequency component and a first low-frequency component, a second target high-frequency component and a second low-frequency component, a third target high-frequency component and a third low-frequency component, and a fourth target high-frequency component and a fourth low-frequency component.

[0026] In the technical solution provided in step S104 of the present invention, if Figure 2 As shown, the first spatial feature is input into the first wavelet transform module of the first stage to obtain the first pair of target high-frequency components and low-frequency components; the second spatial feature is input into the second wavelet transform module of the second stage to obtain the second pair of target high-frequency components and low-frequency components; the third spatial feature is input into the third wavelet transform module of the third stage to obtain the third pair of target high-frequency components and low-frequency components; the fourth spatial feature is input into the fourth wavelet transform module of the fourth stage to obtain the fourth pair of target high-frequency components and low-frequency components.

[0027] Step S105, taking the target high-frequency component and low-frequency component of the fourth spatial feature as the output of the fourth frequency domain feature compensation module, and outputting a fourth high-frequency feature and a fourth low-frequency feature.

[0028] In the technical solution provided in step S105 of the present invention, if Figure 2 As shown, the target high-frequency component and low-frequency component of the fourth spatial feature are used as the output of the fourth frequency domain feature compensation module, and the fourth high-frequency feature and the fourth low-frequency feature are output.

[0029] Step S106, input the fourth high-frequency feature and the fourth low-frequency feature, the third target high-frequency component and the third low-frequency component into the third frequency domain feature compensation module to obtain the third high-frequency feature and the third low-frequency feature, and repeat forward to obtain the second high-frequency feature and the second low-frequency feature of the second frequency domain feature compensation module, and the first high-frequency feature and the first low-frequency feature of the first frequency domain feature compensation module.

[0030] In the technical solution provided in step S106 of the present invention, if Figure 2 As shown, the fourth high-frequency feature and the fourth low-frequency feature, the third target high-frequency component and the third low-frequency component are input into the third frequency domain feature compensation module to obtain the third high-frequency feature and the third low-frequency feature, and this operation is performed on different stages to obtain the second high-frequency feature and the second low-frequency feature of the second frequency domain feature compensation module and the first high-frequency feature and the first low-frequency feature of the first frequency domain feature compensation module.

[0031] Step S107, input the fourth resolution feature into the feature after the dilated spatial convolution pooling pyramid, the fourth spatial feature and the first low-frequency feature into the fourth decoder to obtain the fourth decoded feature, input the fourth decoded feature and the third spatial feature into the third decoder to obtain the third decoded feature, repeat forward to obtain the second decoded feature.

[0032] In the technical solution provided in step S107 of the present invention, if Figure 2 As shown, the fourth resolution feature is input into the feature after the atrous spatial convolution pooling pyramid, the fourth spatial feature and the first low-frequency feature are input into the fourth decoder to obtain the fourth decoded feature, the fourth decoded feature and the third spatial feature are input into the third decoder to obtain the third decoded feature, and the process is repeated in different stages to obtain the second decoded feature.

[0033] Step S108: input the first high-frequency feature, the first spatial feature, and the second decoded feature into a first decoder to obtain a first decoded feature.

[0034] In the technical solution provided in step S108 of the present invention, if Figure 2 As shown, the first high-frequency feature, the first spatial feature and the second decoded feature are input into the first decoder to obtain the first decoded feature.

[0035] Step S109: input the first decoded feature into the convolution layer and the sigmoid function in sequence to obtain the segmentation result of the high-resolution road remote sensing image.

[0036] In the technical solution provided in the above step S109 of the present invention, the first decoded feature is sequentially input into a 1×1 convolutional layer and a sigmoid function to obtain a segmentation result of a high-resolution road remote sensing image.

[0037] The above method of this embodiment is further introduced below.

[0038] As an optional implementation method, step S103, the four different resolution features are input into the corresponding four spatial perception mamba modules to obtain four different spatial features, including: inputting the four different resolution features into the pyramid pooling layer to obtain four pooling features of different sizes; upsampling the four different size pooling features to the same size as the four different resolution features; splicing the upsampled features of the four different size pooling features with the four different resolution features to obtain the spliced ​​features of the four different resolution features; inputting the spliced ​​features of the four different resolution features into the visual state space module to obtain four different spatial features.

[0039] In this embodiment, the four different resolution features are input into the corresponding four spatial perception Mamba modules to obtain the expressions of four different spatial features:

[0040] in, For four different spatial features, is a pooling operation. The value of is 1, 2, 3, 6, which means that the feature map is pooled to the corresponding spatial size 1×1, 2×2, 3×3, 6×6; represents the upsampling operation, represents the operation of concatenating features in the channel dimension, Represents a linear state-space operation.

[0041] As an optional implementation method, step S104, the four different spatial features are input into the corresponding four wavelet transform modules to obtain four pairs of different target high-frequency components and low-frequency components, including: the four different spatial features are passed through four wavelet transform modules to obtain four pairs of three different high-frequency components and one low-frequency component; the four pairs of three different high-frequency components are connected to obtain four pairs of different target high-frequency components.

[0042] As an optional implementation method, step S106, the fourth high-frequency feature and the fourth low-frequency feature, the third target high-frequency component and the third low-frequency component are input into the third frequency domain feature compensation module to obtain the third high-frequency feature and the third low-frequency feature, including: filtering the fourth high-frequency feature through the high-pass filter in the third frequency domain feature compensation module to obtain a fourth high-frequency feature map; filtering the fourth low-frequency feature through the low-pass filter in the third frequency domain feature compensation module to obtain a fourth low-frequency feature map; processing the fourth high-frequency feature map, the third target high-frequency component and the fourth high-frequency feature through the window-based linear model in the third frequency domain feature compensation module, and inputting them into the linear combination layer with the third target high-frequency component to obtain the third high-frequency feature; processing the fourth low-frequency feature map, the third low-frequency component and the fourth low-frequency feature through the window-based linear model in the third frequency domain feature compensation module, and inputting them into the linear combination layer with the third low-frequency component to obtain the third low-frequency feature.

[0043] In this embodiment, after filtering the fourth high-frequency feature through the high-pass filter in the third frequency domain feature compensation module, the expression of the fourth high-frequency feature map is obtained as follows:

[0044] in, ,when hour, is the fourth high-frequency feature map, is the fourth high frequency feature, represents the joint attention composed of the spatial attention mechanism and the channel attention mechanism, Represents the normalization module composed of residual block and batch normalization.

[0045] After filtering the fourth low-frequency feature through the low-pass filter in the third frequency domain feature compensation module, the expression of the fourth low-frequency feature map is obtained as follows:

[0046] in, ,when hour, is the fourth low-frequency feature map, is the fourth low-frequency feature, represents position normalization; Represents the normalization module consisting of residual block and instance normalization.

[0047] After the fourth high-frequency feature map, the third target high-frequency component and the fourth high-frequency feature are processed by the window-based linear model in the third frequency domain feature compensation module, they are input into the linear combination layer with the third target high-frequency component, and the expression of the third high-frequency feature is obtained as follows:

[0048] in, , PFFC It is a progressive frequency feature compensation method (PFFC for short). hour, is the fourth high frequency feature map, is the third target high frequency component, It is the fourth high frequency feature.

[0049] After the fourth low-frequency feature map, the third low-frequency component and the fourth low-frequency feature are processed by the window-based linear model in the third frequency domain feature compensation module, they are input into the linear combination layer with the third low-frequency component, and the expression of the third low-frequency feature is obtained as follows:

[0050] in, , PFFC It is a progressive frequency feature compensation method (PFFC for short). hour, is the fourth low-frequency characteristic map, is the third low frequency component, It is the fourth low-frequency feature.

[0051] As an optional implementation manner, in step S102, the backbone networks of the four encoders with different resolutions are all Mamba.

[0052] As an optional implementation manner, in step S107, each decoder is composed of a strip convolution and upsampling module.

[0053] In an embodiment of the present invention, a high-resolution road remote sensing image is acquired; the high-resolution road remote sensing image is sequentially input into four encoders with different resolutions to obtain four different resolution features corresponding to the high-resolution road remote sensing image, wherein the four different resolution features are respectively a first resolution feature, a second resolution feature, a third resolution feature and a fourth resolution feature; the four different resolution features are input into corresponding four spatial perception Mamba modules to obtain four different spatial features, wherein the four different spatial features are respectively a first spatial feature, a second spatial feature, a third spatial feature and a fourth spatial feature; the four different spatial features are input into corresponding four wavelet transform modules to obtain four pairs of different target high-frequency components and low-frequency components, wherein the four pairs of different target high-frequency components and low-frequency components are respectively a first target high-frequency component and a first low-frequency component, a second target high-frequency component and a second low-frequency component, a third target high-frequency component and a third low-frequency component, and a fourth target high-frequency component and a fourth low-frequency component; the target high-frequency component and low-frequency component of the fourth spatial feature are used as the output of the fourth frequency domain feature compensation module, and the fourth high-frequency feature and the fourth low-frequency feature are output; the fourth high-frequency feature and the fourth low-frequency feature are respectively input into the corresponding four wavelet transform modules. The frequency feature, the third target high-frequency component and the third low-frequency component are input into the third frequency domain feature compensation module to obtain the third high-frequency feature and the third low-frequency feature, and the process is repeated forward to obtain the second high-frequency feature and the second low-frequency feature of the second frequency domain feature compensation module, and the first high-frequency feature and the first low-frequency feature of the first frequency domain feature compensation module respectively; the fourth resolution feature is input into the feature after the void space convolution pooling pyramid, the fourth spatial feature and the first low-frequency feature are input into the fourth decoder to obtain the fourth decoded feature, the fourth decoded feature and the third spatial feature are input into the third decoder to obtain the third decoded feature, and the process is repeated forward to obtain the second decoded feature; the first high-frequency feature, the first spatial feature and the second decoded feature are input into the first decoder to obtain the first decoded feature; the first decoded feature is sequentially input into the convolution layer and the sigmoid function to obtain the segmentation result of the high-resolution road remote sensing image, which solves the technical problem of inaccurate road segmentation caused by the similarity and occlusion problem between the road and the background in the high-resolution road remote sensing image in the prior art, and achieves the technical effect of accurate road segmentation through the new network framework.

[0054] The serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.

[0055] In the above embodiments of the present invention, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0056] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the device embodiments described above are only schematic. For example, the division of units can be a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of units or modules, which can be electrical or other forms.

[0057] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed over multiple units. Some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0058] In addition, each functional unit in each embodiment of the present invention may be integrated into a first processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0059] The above are only preferred embodiments of the present invention. It should be pointed out that, for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A selective state space model road segmentation method based on frequency domain feature compensation, characterized in that: include: Acquire high-resolution road remote sensing images; The high-resolution road remote sensing image is sequentially input into four encoders with different resolutions to obtain four different resolution features corresponding to the high-resolution road remote sensing image, wherein the four different resolution features are respectively a first resolution feature, a second resolution feature, a third resolution feature and a fourth resolution feature; Inputting four different resolution features into corresponding four spatial perception Mamba modules to obtain four different spatial features, wherein the four different spatial features are respectively a first spatial feature, a second spatial feature, a third spatial feature and a fourth spatial feature; Inputting four different spatial features into corresponding four wavelet transform modules, four different pairs of target high-frequency components and low-frequency components are obtained, wherein the four different pairs of target high-frequency components and low-frequency components are respectively a first target high-frequency component and a first low-frequency component, a second target high-frequency component and a second low-frequency component, a third target high-frequency component and a third low-frequency component, and a fourth target high-frequency component and a fourth low-frequency component; Taking the target high-frequency component and low-frequency component of the fourth spatial feature as the output of the fourth frequency domain feature compensation module, outputting a fourth high-frequency feature and a fourth low-frequency feature; Input the fourth high-frequency feature and the fourth low-frequency feature, the third target high-frequency component and the third low-frequency component into the third frequency domain feature compensation module to obtain the third high-frequency feature and the third low-frequency feature, and repeat forward to obtain the second high-frequency feature and the second low-frequency feature of the second frequency domain feature compensation module, and the first high-frequency feature and the first low-frequency feature of the first frequency domain feature compensation module respectively; Input the fourth resolution feature into the feature after the dilated spatial convolution pooling pyramid, the fourth spatial feature and the first low-frequency feature into the fourth decoder to obtain the fourth decoded feature, input the fourth decoded feature and the third spatial feature into the third decoder to obtain the third decoded feature, repeat forward to obtain the second decoded feature; Inputting the first high-frequency feature, the first spatial feature, and the second decoded feature into a first decoder to obtain a first decoded feature; The first decoded features are sequentially input into the convolutional layer and the sigmoid function to obtain the segmentation result of the high-resolution road remote sensing image.

2. The method according to claim 1, characterized in that The four different resolution features are input into the corresponding four spatial perception Mamba modules to obtain four different spatial features, including: Input four features of different resolutions into the pyramid pooling layer to obtain four pooling features of different sizes; Upsample the pooled features of four different sizes to the same size as the features of four different resolutions; The upsampled features of the four pooling features of different sizes are concatenated with the four features of different resolutions to obtain the concatenated features of the four features of different resolutions; The concatenated features of four features with different resolutions are input into the visual state space module to obtain four different spatial features.

3. The method according to claim 1, characterized in that The four different spatial features are input into the corresponding four wavelet transform modules to obtain four different pairs of target high-frequency components and low-frequency components, including: The four different spatial features are passed through four wavelet transform modules to obtain four different pairs of three high-frequency components and one low-frequency component; Four pairs of three different high-frequency components are connected to obtain four different pairs of target high-frequency components.

4. The method according to claim 1, characterized in that The fourth high frequency feature and the fourth low frequency feature, the third target high frequency component and the third low frequency component are input into the third frequency domain feature compensation module to obtain the third high frequency feature and the third low frequency feature, including: After filtering the fourth high-frequency feature through a high-pass filter in the third frequency domain feature compensation module, a fourth high-frequency feature map is obtained; After filtering the fourth low-frequency feature through a low-pass filter in the third frequency domain feature compensation module, a fourth low-frequency feature map is obtained; After processing the fourth high-frequency feature map, the third target high-frequency component and the fourth high-frequency feature through the window-based linear model in the third frequency domain feature compensation module, the fourth high-frequency feature and the third target high-frequency component are input into the linear combination layer to obtain the third high-frequency feature; After the fourth low-frequency feature map, the third low-frequency component and the fourth low-frequency feature are processed by the window-based linear model in the third frequency domain feature compensation module, they are input into the linear combination layer with the third low-frequency component to obtain the third low-frequency feature.

5. The method according to claim 1, characterized in that The backbone networks of the four encoders with different resolutions are all Mamba.

6. The method according to claim 1, characterized in that Each decoder is composed of strip convolution and upsampling modules.

7. A computer system, characterized in that include: One or more processors, and a computer-readable storage medium for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors are enabled to implement the method of claim 1.

8. A computer-readable storage medium, characterized in that Computer executable instructions are stored, and when the instructions are executed, they are used to implement the method of claim 1.

9. A computer program product, characterized in that The invention comprises computer executable instructions, which are used to implement the method of claim 1 when being executed.

Images