Remote sensing image road segmentation method, system and equipment based on mixed scale feature extraction and dynamic intelligent kernel feature enhancement, and medium

Abstract

The invention relates to the technical field of remote sensing image processing, and discloses a remote sensing image road segmentation method, system and device based on mixed scale feature extraction and dynamic intelligent kernel feature enhancement, and a medium, and the method comprises the steps: extracting different scale road information of a remote sensing image based on a mixed scale feature extraction unit, and obtaining a remote sensing image feature map; performing multi-scale feature and global context information extraction on the remote sensing image feature map through a dynamic intelligent kernel feature enhancement unit to obtain a feature enhancement map; and performing road segmentation on the feature enhancement graph based on Mama enhancement and a Diffusion model, and outputting a road segmentation result. Not only is the accuracy of road segmentation improved, but also the problems in the prior art that multi-scale information processing is insufficient and the segmentation precision is reduced in a complex background scene are effectively solved, and a more efficient and accurate solution is provided for an intelligent traffic system.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing image processing, and more specifically, to a remote sensing image road segmentation method, system, device and medium based on hybrid scale feature extraction and dynamic intelligent core feature enhancement. Background Art

[0002] With the acceleration of the urbanization process, the construction of intelligent transportation systems has become an important direction for improving traffic efficiency and ensuring traffic safety. Road segmentation, as one of the basic tasks in intelligent transportation systems, directly affects the accuracy of subsequent technologies such as road recognition and vehicle navigation. Especially in complex environments, how to accurately segment roads, especially in scenes with blurred details or complex backgrounds, has become a research difficulty.

[0003] At present, road segmentation methods based on deep learning have achieved certain results. Traditional road segmentation methods mainly rely on image processing technologies such as edge detection and color segmentation. These methods can achieve good results in cases with simple backgrounds or clear roads, but in complex backgrounds, especially in harsh environments such as night, rain, and snow, the performance of these methods is greatly reduced. To address this challenge, methods based on deep neural networks (DNNs) have been gradually proposed and applied to road segmentation tasks. Convolutional neural networks (CNNs), as an important architecture among them, have shown excellent performance in many visual tasks through their powerful feature extraction capabilities. However, existing deep learning methods generally have two problems: one is the inability to effectively process road information at different scales; the other is that in scenes with complex backgrounds or relatively blurred road details, the segmentation accuracy of the model is significantly affected.

[0004] To solve these two problems, some methods in the existing technologies attempt to enhance the robustness of the model through multi-scale feature fusion. For example, some networks based on the U-Net structure improve the segmentation effect by gradually fusing features at different levels, but these methods still cannot fully play their roles in practical applications, especially for the precise capture of road details. And some other methods enhance the expression of important features by introducing attention mechanisms, but these methods usually rely on large-scale data training and still do not perform well in dealing with complex backgrounds.

[0005] Therefore, how to provide a remote sensing image road segmentation method based on hybrid scale feature extraction and dynamic intelligent core feature enhancement is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0006] In view of this, the present invention provides a remote sensing image road segmentation method, system, device and medium based on hybrid scale feature extraction and dynamic intelligent core feature enhancement, aiming to improve the accuracy and robustness of road segmentation in remote sensing images by improving the feature extraction and enhancement modules, especially for road recognition in complex backgrounds, and reducing the computational resource consumption of model training and deployment. Through multi-scale feature extraction and adaptive dynamic enhancement, the present invention can effectively capture road information in the image and has strong adaptability to different environments.

[0007] To achieve the above object, the present invention adopts the following technical solutions:

[0008] A remote sensing image road segmentation method based on hybrid scale feature extraction and dynamic intelligent core feature enhancement, comprising:

[0009] Extracting different scale road information of a remote sensing image based on a hybrid scale feature extraction unit to obtain a remote sensing image feature map;

[0010] Performing multi-scale feature and global context information extraction on the remote sensing image feature map through a dynamic intelligent core feature enhancement unit to obtain a feature enhanced map;

[0011] Performing road segmentation on the feature enhanced map based on the Mamba enhancement and Diffusion model, and outputting a road segmentation result.

[0012] Preferably, extracting different scale road information of a remote sensing image based on a hybrid scale feature extraction unit to obtain a remote sensing image feature map specifically includes:

[0013] Step 1.1: Preprocessing the remote sensing image;

[0014] Step 1.2: Inputting the preprocessed remote sensing image into an initial convolutional layer to obtain an initial feature map f o ;

[0015] Step 1.3: After passing through the normalization layer, the initial feature map f o is respectively subjected to upper branch and lower branch feature extraction, and the features of the upper branch and the lower branch are spliced to correspondingly obtain a feature map f 1 and a feature map f 2 ;

[0016] Step 1.4: The feature map f 1 successively passes through a 3×3 depthwise separable convolutional layer and a ReLU activation function to obtain a feature map f 1' ;

[0017] The feature map f 2 successively passes through a 5×5 depthwise separable convolutional layer and a ReLU activation function to obtain a feature map f 2′ ;

[0018] Step 1.5: Concatenate the feature map f 1' and the feature map f 2' to obtain a fused feature map f + ;

[0019] Step 1.6: After performing windmill-shaped convolution on the fused feature map f + and add it to the initial feature map f o to obtain the remote sensing image feature map

[0020] Preferably, the multi-scale features and global context information of the remote sensing image feature map are extracted through a dynamic intelligent core feature enhancement unit to obtain a feature enhancement map, specifically including:

[0021] Step 2.1: Perform linear projection on the remote sensing image feature map and input it into depthwise separable convolutional layers of different sizes to obtain two feature maps and the feature map

[0022] Step 2.2: The feature map successively passes through an average pooling layer, a dynamic deformable convolutional layer, and a sigmoid activation function to obtain a weighted feature map A 5×5 ;

[0023] The feature map successively passes through a max pooling layer, a full-dimensional dynamic convolutional layer, and a sigmoid activation function to obtain a weighted feature map A 7×7 ;

[0024] Step 2.3: Concatenate the feature map and the feature map Multiply the concatenated feature map with the weighted feature map A 5×5 and the weighted feature map A 7×7 respectively, and fuse the two corresponding weighted feature maps to obtain a feature map

[0025] Step 2.4: The feature map passes through an average global pooling layer, a convolutional layer, and a sigmoid activation function respectively to obtain weights

[0026] Step 2.5: Perform a pointwise multiplication operation on the weights and the feature map and perform a skip connection on the features after pointwise multiplication with the features in the remote sensing image feature map to obtain a feature enhancement map

[0027] Preferably, road segmentation is performed on the feature enhancement map based on the Mamba enhancement and Diffusion models, and the road segmentation result is output, specifically including:

[0028] Step 3.1: After passing through the super-resolution module, the feature enhancement map expands the resolution of the feature map through upsampling, and at the same time enhances the detail area through Mamba enhancement;

[0029] Step 3.2: After being processed in Step 3.1, remove the noise through the Diffusion model;

[0030] Step 3.3: After being processed in Step 3.2, perform classification through the convolutional layer and output the road segmentation result.

[0031] Preferably, Step 1.3 specifically includes:

[0032] When extracting features in the upper branch, the initial feature map f o passes through a 1×1 convolutional layer, a 3×3 depthwise separable convolutional layer, and a ReLU activation function in sequence to obtain the feature f up ;

[0033] When extracting features in the lower branch, the initial feature map f o passes through a 1×1 convolutional layer, a 5×5 depthwise separable convolutional layer, and a ReLU activation function in sequence to obtain the feature f down ;

[0034] Concatenate the feature f up and the feature f down to correspondingly obtain the feature map f 1 and the feature map f 2 , and the calculation formula is:

[0035] f 1 = Concat(f up , f down ), f 2 = Concat(f down , f up )

[0036] where Concat represents the concatenation operation of the feature maps.

[0037] Preferably, the model loss function is:

[0038] L total = L1 + α·L2 + β·L3

[0039]

[0040] Among them, α and β are hyperparameters for balancing various losses, L1 is the cross-entropy loss, L2 is the Dice loss, L3 is the adaptive regularization loss, and y i is the actual label, is the predicted output of the model, λ is the regularization strength, ||·||2 represents the L2 norm, and i represents the pixel points in the data sample.

[0041] A remote sensing image road segmentation system based on hybrid scale feature extraction and dynamic intelligent core feature enhancement, comprising:

[0042] A hybrid scale feature extraction unit: used to extract road information at different scales of the remote sensing image to obtain a remote sensing image feature map;

[0043] A dynamic intelligent core feature enhancement unit: used to extract multi-scale features and global context information from the remote sensing image feature map to obtain a feature enhanced map;

[0044] A road segmentation unit: used to perform road segmentation on the feature enhanced map based on Mamba enhancement and the Diffusion model, and output the road segmentation result.

[0045] A computer device, comprising: a memory and a processor, wherein a computer program that can run on the processor is stored in the memory, and when the processor executes the computer program, a remote sensing image road segmentation method based on hybrid scale feature extraction and dynamic intelligent core feature enhancement is implemented.

[0046] A computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, a remote sensing image road segmentation method based on hybrid scale feature extraction and dynamic intelligent core feature enhancement is implemented.

[0047] It can be seen from the above technical solutions that, compared with the prior art, the present invention discloses a remote sensing image road segmentation method, system, device and medium based on hybrid scale feature extraction and dynamic intelligent core feature enhancement, having the following advantages:

[0048] 1) The hybrid scale feature extraction module can extract road features at multiple different scales by fusing multi-scale information, significantly improving the model's adaptability to different image scales. In practical applications, the size of the image, the width, shape, and complexity of the road often vary. Traditional single-scale feature extraction methods are prone to losing important information. The hybrid scale feature extraction module performs multi-scale feature extraction on the input image by using convolution kernels of different scales (such as 1×1, 3×3, 5×5, etc.), and further integrates these feature information of different scales to ensure accurate road segmentation regardless of the specific scale of the road in the image. The design of this module can better capture the road framework from large scales to road details in the image, improving the model's performance on images of different sizes, especially for road detection and segmentation tasks in complex scenarios.

[0049] 2) Under changing environmental conditions, the background of the image may be very complex, and the details of the road may become blurred or unclear due to various factors. Traditional methods often have difficulty accurately distinguishing the road from the background. The dynamic intelligent core feature enhancement module is designed specifically for this problem. It improves the expressive ability of features by adaptively selecting the most representative and discriminative features in the image, thereby improving the segmentation accuracy. This module uses self-attention mechanisms or similar adaptive weighting methods to dynamically adjust their contributions in the model according to the importance of different regions or features, ensuring the enhancement of key features in subsequent calculations. For example, in complex background areas, it can automatically amplify road-related features while suppressing irrelevant background information. In the case of blurred road details, this module can automatically enhance the capture of road contours, preventing over-smoothing and resulting in detail loss, thus effectively improving the segmentation accuracy, especially in cases where the background is complex or road details are unclear. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only the embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained according to the provided drawings without creative efforts.

[0051] Figure 1 It is a flowchart of a remote sensing image road segmentation method based on hybrid scale feature extraction and dynamic intelligent core feature enhancement provided by the present invention.

[0052] Figure 2 It is a processing flowchart of the hybrid scale feature extraction unit provided by the present invention.

[0053] Figure 3The processing flow chart of the dynamic intelligent core feature enhancement unit provided by the present invention.

[0054] Figure 4 The structural block diagram of the remote sensing image road segmentation system based on hybrid scale feature extraction and dynamic intelligent core feature enhancement provided by the present invention. Specific implementation manners

[0055] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0056] The embodiment of the present invention discloses a method for segmenting roads in remote sensing images with hybrid scale feature extraction and dynamic intelligent core feature enhancement, as Figure 1 shown, including:

[0057] Extract different scale road information of the remote sensing image based on the hybrid scale feature extraction unit to obtain a remote sensing image feature map;

[0058] Extract multi-scale features and global context information from the remote sensing image feature map through the dynamic intelligent core feature enhancement unit to obtain a feature enhanced map;

[0059] Perform road segmentation on the feature enhanced map based on the Mamba enhancement and Diffusion model, and output the road segmentation result.

[0060] In this embodiment, the specific processing process of the hybrid scale feature extraction unit is as Figure 2 shown.

[0061] Step 1.1: Preprocess the input remote sensing image. First, the remote sensing image usually contains noise and different lighting conditions, so preprocessing is required to improve the performance of the model. Normalization: Adjust the pixel value range to between [0, 1] to ensure the stability of the gradient during network training. Denoising: Remove the noise in the remote sensing image through a filtering operation. Image enhancement: Adjust the brightness, contrast, etc. of the image to increase the diversity of training samples. The preprocessed image will be used as the input of the neural network and represented as a three-dimensional matrix: where H is the height of the image, W is the width of the image, and C is the number of channels of the image.

[0062] Step 1.2: Feature extraction of the initial convolutional layer. After preprocessing, the image The initial convolutional layer of the incoming network is used to extract preliminary feature information. The role of this convolutional layer is to extract low-level features of the image (such as edges, textures, etc.) by learning convolutional kernels. The output of the initial convolutional layer is the initial feature map f o 。

[0063] Step 1.3: Feature extraction of the dual-branch network. Next, the initial feature map f extracted by the initial convolutional layer o will be passed to the normalization layer. The role of the normalization layer is to standardize the features, making them have zero mean and unit variance, and further improving the convergence speed and robustness of the network. After the normalization layer, the feature map will enter the dual-branch structure of the network and undergo feature extraction through the upper branch and the lower branch respectively. The design of the two branches aims to extract features at different scales to enhance the network's learning ability for road structures in remote sensing images.

[0064] Upper branch: The upper branch first adjusts the number of channels of the input features through a 1×1 convolutional layer, and then performs local feature extraction through a 3×3 depthwise separable convolutional layer, and obtains the feature f through the ReLU activation function up 。

[0065] Lower branch: The lower branch adjusts the number of channels through a 1×1 convolutional layer, and then uses a 5×5 depthwise separable convolutional layer to perform feature extraction over a larger range, and also obtains the feature f through the ReLU activation function down 。

[0066] Next, the features of the upper branch and the lower branch are concatenated. First, cross-concatenation is performed to obtain two new feature maps f 1 and f 2 。

[0067] f 1 =Concat(f up ,f down ), f 2 =Concat(f down ,f up )

[0068] Here, Concat represents the concatenation operation of the feature maps.

[0069] Step 1.4: Further feature extraction. The concatenated f 1 and f 2 are respectively fed into the upper branch and the lower branch for further feature extraction.

[0070] Upper branch: Process the feature map f 1 through a 3×3 depthwise separable convolutional layer and the ReLU activation function to obtain a new feature map f 1′ 。

[0071] Lower branch: Process the feature f 2 through a 5×5 depthwise separable convolutional layer and a ReLU activation function to obtain a new feature map f 2′ .

[0072] Step 1.5: Feature fusion and output. Concatenate f 1′ and f 2’ to obtain the final fused feature map f + . This fused feature map f + contains road information from different scales, enhancing the representation ability of the road segmentation task.

[0073] Step 1.6: The fused feature map f + is subjected to a pinwheel-shaped convolution (PSConv) operation and added to the initial feature map f o to obtain the final remote sensing image feature map This step combines local and global features to obtain richer and more accurate road segmentation information, laying a foundation for subsequent segmentation tasks.

[0074] The hybrid scale feature extraction unit of the present invention can capture road information at different scales through a multi-scale feature extraction method. By combining convolutional kernels of different sizes, large-scale frame features and small-scale detail features in the image are extracted respectively, ensuring that the model can still maintain good performance when processing images of different sizes and avoiding the neglect of detail information at certain scales by traditional methods. Through the above process, the network can effectively extract road features in remote sensing images, providing reliable input for the road segmentation task.

[0075] In this embodiment, the specific processing process of the dynamic intelligent core feature enhancement unit is as Figure 3 shown. This step aims to construct a dynamic intelligent core feature enhancement unit to effectively extract multi-scale features and global context information through this module, enhancing the performance of remote sensing image road segmentation.

[0076] Step 2.1: Feature input and linear projection. The remote sensing image feature map undergoes linear projection to reduce its number of channels to a smaller dimension. This step is completed through a 1×1 convolutional layer operation, reducing the computational complexity and preparing for subsequent feature extraction. This operation reduces the number of channels of the feature map, reduces the computational complexity, and at the same time maintains the integrity of the feature information, helping to improve the computational efficiency. Through this optimization, subsequent convolutional operations can be performed at a lower computational cost without sacrificing the performance of the segmentation task.

[0077] To effectively extract features at different scales, it is input into depthwise separable convolutional layers of two different sizes: 5×5 and 7×7 after linear projection. These two convolutional layers generate two feature maps respectively: and Depthwise separable convolution can significantly reduce the computational load while capturing features at different scales and adapting to the different sizes of targets such as roads in remote sensing images. By using multiple large-sized convolutional kernels, the model can gradually expand the receptive field, better capture long-range dependencies, and understand global context information.

[0078] Step 2.2: Dynamic feature selection and enhancement. Next, and are dynamically selected and enhanced. For first, dimensionality reduction is performed through an average pooling layer, then it is input into a dynamic deformable convolutional layer for adaptive feature selection, and finally, a weighted feature map A 5×5 :

[0079]

[0080] where σ represents the sigmoid activation function, DCNv4 represents the dynamic deformable convolution, and AVP represents the average pooling layer. This dynamic adjustment mechanism adaptively highlights the most important channel and spatial features according to the global information of the input features, thereby improving the robustness and adaptability of the model and enabling the features to be better separated from the complex background.

[0081] For dimensionality reduction is performed through a max pooling layer, then it is input into a full-dimensional dynamic convolutional layer, and finally, a weighted feature map A 7×7 :

[0082]

[0083] where ODConv represents the full-dimensional dynamic convolutional layer and MAP represents the max pooling layer. This dynamic selection mechanism can guide the selection and weighting of features according to the global information, further improving the segmentation accuracy and ensuring that important features are not ignored, especially when processing remote sensing images with multi-scale features.

[0084] Step 2.3: Multi-scale feature fusion. The feature maps from different scales and are concatenated. By concatenating feature maps of different scales, the model can retain the key information captured from each scale and ensure that this information can be effectively fused in subsequent steps. Subsequently, the concatenated feature map is combined with the dynamically selected weighted feature map A 5×5 and A 7×7Multiply to obtain the weighted feature map. Finally, these two feature maps are added and fused to obtain the feature map This dynamic fusion mechanism ensures that information from different scales can be fully combined while retaining the most critical features, significantly improving the detection ability of targets such as roads in remote sensing images.

[0085] Step 2.4: Global pooling and feature enhancement. The feature map Undergoes global information compression through the average global pooling layer, and then the final weight is obtained through the convolutional layer and the sigmoid activation function

[0086]

[0087] Among them, GAP represents global average pooling. The weight Weights the feature map to highlight the parts most helpful for segmentation.

[0088] Step 2.5: Multiply the weight with to perform a pointwise multiplication operation. This integration of global information not only improves the segmentation accuracy but also enhances the generalization ability of the model in different backgrounds. Connect the feature after pointwise multiplication with the remote sensing image feature map through a skip connection to finally obtain the feature enhancement map

[0089]

[0090] Among them, represents pointwise multiplication, represents the addition and fusion operation. Through the skip connection, the model can utilize the comprehensive information of the original features and the enhanced features to further improve the accuracy and effect of remote sensing image road segmentation.

[0091] The dynamic intelligent core feature enhancement unit of the present invention ensures that the most critical features are enhanced in complex backgrounds and blurred road scenes by adaptively adjusting the weights of the features. During the segmentation process, this module can intelligently identify which features are the most important and dynamically adjust their influence on the segmentation result, thereby effectively improving the segmentation accuracy. Especially in cases where the background is complex or the road details are unclear, it can prevent the loss of details caused by over-smoothing of the model.

[0092] In this embodiment, the final segmentation result and the training loss function are obtained based on Mamba enhancement and the Diffusion model.

[0093] Step 3.1: Super-resolution and Mamba enhancement. To improve the accuracy of remote sensing image segmentation, especially in the restoration and enhancement of detail areas, we will It is input into a super-resolution module, which helps to restore the detailed features of the image by improving the spatial resolution of the image. The output of the super-resolution module expands the resolution of the feature map through upsampling, refining the detailed parts of the image, which is particularly important in the segmentation of edges and roads. In this process, we incorporate the Mamba model (a diffusion-based generative model). The Mamba enhancement module can further strengthen the detailed areas through generative adversarial learning, especially for repairing blurred or unclear edges. The introduction of the Mamba model further optimizes the image details, enhances the high-frequency information of the image, and makes the segmentation results more accurate.

[0094] Step 3.2: Diffusion model and global context modeling. To further enhance the global context modeling ability, we introduce the Diffusion Model. This model restores the details of the image from the noise by gradually guiding the removal of noise and simulates the latent structure of the image to capture more global information. The output generated by the Diffusion model can help the model understand the global background in remote sensing images and enhance the distinction between the background and the foreground. This model helps to handle the subtle differences between complex backgrounds and roads and enhances the adaptability of the model in different environments.

[0095] Step 3.3: Generate the final segmentation result. After being processed by super-resolution, Mamba enhancement, and the Diffusion model, the obtained feature map is input into a convolutional layer for final classification to generate the class label for each pixel point. The final segmentation result represents the class to which each pixel point belongs.

[0096] To train this model, a composite loss function is designed, which combines cross-entropy loss, Dice loss, and a new adaptive regularization loss to ensure that the model has good robustness and adaptability when processing remote sensing images. Cross-entropy loss: Cross-entropy loss is used to measure the difference between the model output and the true label, especially suitable for pixel-level classification tasks. In the road segmentation of remote sensing images, cross-entropy loss can help the model effectively distinguish roads and backgrounds:

[0097]

[0098] where y i is the actual label, is the predicted output of the model. Dice loss: Dice loss is used to measure the degree of overlap between the prediction result and the actual label, especially suitable for dealing with imbalanced data (such as road segmentation tasks). It can improve the segmentation accuracy of the model for small objects (such as small roads):

[0099]

[0100] Adaptive regularization loss: The adaptive regularization loss is used to enhance the generalization ability of the model and ensure that the model can still maintain high segmentation performance in complex scenarios. This loss function helps the model adaptively select useful features by dynamically adjusting parameters:

[0101]

[0102] Among them, λ is the regularization strength, and ||·||2 represents the L2 norm, which is used to measure the difference between features. Final loss function: The above losses are weighted and summed to obtain the final training loss function:

[0103] L total = L1 + α·L2 + β·L3

[0104] Among them, α and β are hyperparameters used to balance the losses.

[0105] An embodiment of the present invention provides a remote sensing image road segmentation system based on hybrid-scale feature extraction and dynamic intelligent core feature enhancement, as Figure 4 shown, including:

[0106] Hybrid-scale feature extraction unit: used to extract road information at different scales of the remote sensing image to obtain a remote sensing image feature map;

[0107] Dynamic intelligent core feature enhancement unit: used to extract multi-scale features and global context information from the remote sensing image feature map to obtain a feature enhancement map;

[0108] Road segmentation unit: used to perform road segmentation on the feature enhancement map based on Mamba enhancement and Diffusion model, and output the road segmentation result.

[0109] The specific implementation processes of the units of the system of the present invention are the same as those in the method part, and will not be elaborated here.

[0110] This embodiment provides a computer device, including: a memory and a processor. A computer program that can run on the processor is stored in the memory. When the processor executes the computer program, a remote sensing image road segmentation method based on hybrid-scale feature extraction and dynamic intelligent core feature enhancement is implemented.

[0111] This embodiment provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, a remote sensing image road segmentation method based on hybrid-scale feature extraction and dynamic intelligent core feature enhancement is implemented.

[0112] Those of ordinary skill in the art can understand that all or part of the steps of implementing the above method embodiments can be completed by hardware related to program instructions. The foregoing program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps including the above method embodiments. The foregoing storage medium includes various media that can store program codes, such as removable storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical discs.

[0113] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. For the same or similar parts among the various embodiments, reference can be made to each other. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple. For the relevant parts, reference can be made to the description in the method part.

[0114] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be obvious to those skilled in the art. The general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but will be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A remote sensing image road segmentation method based on mixed scale feature extraction and dynamic intelligence core feature enhancement, characterized in that: include: Based on the mixed-scale feature extraction unit, the road information of different scales of the remote sensing image is extracted to obtain the remote sensing image feature map; The dynamic intelligent core feature enhancement unit is used to extract multi-scale features and global context information from the remote sensing image feature map to obtain a feature enhancement map. Based on Mamba enhancement and Diffusion model, the feature enhancement map is segmented and the road segmentation result is output.

2. The method for remote sensing image road segmentation based on hybrid scale feature extraction and dynamic intelligence core feature enhancement according to claim 1, characterized in that: Based on the mixed scale feature extraction unit, the road information of different scales of the remote sensing image is extracted to obtain the remote sensing image feature map, which specifically includes: Step 1.1: Preprocess the remote sensing image; Step 1.2: Input the preprocessed remote sensing image into the initial convolution layer to obtain the initial feature map f o ; Step 1.3: Initial feature map f o After passing through the normalization layer, the upper branch and lower branch features are extracted respectively, and the features of the upper branch and the lower branch are spliced ​​to obtain the corresponding feature map f 1 And the feature map f 2 ; Step 1.4: Feature map f 1 Through the 3×3 depth separable convolution layer and ReLU activation function in sequence, the feature map f is obtained 1′ ; Feature map f 2 The feature map f is obtained by sequentially passing through a 5×5 depth-separable convolutional layer and a ReLU activation function. 2′ ; Step 1.5: Transform the feature map f 1′ And the feature map f 2′ Splice to get the fusion feature map f + ; Step 1.6: Fusion feature map f + After the windmill convolution and the initial feature map f o Add them together to get the remote sensing image feature map 3. The method for remote sensing image road segmentation based on hybrid scale feature extraction and dynamic intelligence core feature enhancement according to claim 1, characterized in that: The dynamic intelligent core feature enhancement unit extracts multi-scale features and global context information from the remote sensing image feature map to obtain a feature enhancement map, which specifically includes: Step 2.1: Linearly project the remote sensing image feature map and input it into depth-separable convolutional layers of different sizes to obtain two feature maps and feature map Step 2.2: Feature Map The weighted feature map is obtained by sequentially passing through the average pooling layer, the dynamic deformable convolution layer and the sigmoid activation function. Feature Map The weighted feature map A is obtained by sequentially passing through the maximum pooling layer, the full-dimensional dynamic convolution layer and the sigmoid activation function. 7×7 ; Step 2.3: Feature map and feature map Splice and combine the spliced ​​feature maps with the weighted feature map A 5×5 And the weighted feature map A 7×7 Multiply them and fuse the two corresponding weighted feature maps to get the feature map Step 2.4: Feature map The weights are obtained by averaging the global pooling layer, convolution layer and sigmoid activation function respectively. Step 2.5: Weight and feature map Perform point-by-point multiplication and compare the features after point-by-point multiplication with the remote sensing image feature map The features in the graph are skipped to obtain the feature enhancement graph 4. The method for remote sensing image road segmentation based on hybrid scale feature extraction and dynamic intelligence core feature enhancement according to claim 1, characterized in that: Based on the Mamba enhancement and Diffusion model, the feature enhancement map is segmented and the road segmentation results are output, including: Step 3.1: After the feature enhancement map passes through the super-resolution module, the resolution of the feature map is expanded by upsampling, and the detail area is enhanced by Mamba enhancement; Step 3.2: After processing in step 3.1, remove the noise through the Diffusion model; Step 3.3: After processing in step 3.2, the data is classified through the convolutional layer and the road segmentation result is output.

5. The method for remote sensing image road segmentation based on hybrid scale feature extraction and dynamic intelligence core feature enhancement according to claim 2, characterized in that: Step 1.3 specifically includes: When extracting the upper branch features, the initial feature map f o After passing through the 1×1 convolution layer, the 3×3 depth-separable convolution layer and the ReLU activation function, the feature f is obtained. up ; When extracting features from the lower branch, the initial feature map f o After passing through the 1×1 convolution layer, the 5×5 depth-separable convolution layer and the ReLU activation function, the feature f is obtained. down ; The feature f up and feature f down Splice and get the corresponding feature map f 1 And the feature map f 2 , the calculation formula is: f 1 =Concat(f up ,f down ),f 2 =Concat(f down ,f up ) Among them, Concat represents the concatenation operation of feature maps.

6. The method for remote sensing image road segmentation based on hybrid scale feature extraction and dynamic intelligence core feature enhancement according to claim 1, characterized in that: The model loss function is: Among them, α and β are hyperparameters used to balance various losses, L1 is the cross entropy loss, L2 is the Dice loss, L3 is the adaptive regularization loss, and y i is the actual label, is the predicted output of the model, λ is the regularization strength, ||·||2 represents the L2 norm, and i represents the pixel in the data sample.

7. A remote sensing image road segmentation system based on mixed scale feature extraction and dynamic intelligence core feature enhancement is characterized by: include: Mixed-scale feature extraction unit: used to extract road information of different scales from remote sensing images to obtain remote sensing image feature maps; Dynamic intelligent core feature enhancement unit: used to extract multi-scale features and global context information from remote sensing image feature maps to obtain feature enhancement maps; Road segmentation unit: used to perform road segmentation on the feature enhancement map based on Mamba enhancement and Diffusion model, and output the road segmentation result.

8. A computer device, characterized in that: include: A memory and a processor, wherein the memory stores a computer program that can be run on the processor, and when the processor executes the computer program, the method according to any one of claims 1 to 6 is implemented.

9. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, which, when executed by a processor, implements the method described in any one of claims 1 to 6.

Images