A remote sensing image road segmentation method combining road detail feature extraction and context information perception

Abstract

This invention provides a remote sensing image road segmentation method that combines road detail feature extraction and context information perception, belonging to the field of road segmentation. It solves the problems of occlusion and difficulty in distinguishing road levels in current road segmentation networks. The method includes the following steps: dataset construction, model construction, model training, and result prediction. This invention designs a road detail feature extraction module and a context information perception module. The road detail feature extraction module effectively compensates for the loss of spatial information and edge details caused by continuous downsampling through the interactive transmission of multi-level feature information, enhancing the model's understanding of complex scenes and multi-level road structures. The context information perception module realizes the dynamic fusion of long-distance and local multi-scale context information, helping to suppress background interference, occlusion effects, and confusion caused by similar road structures, ensuring the integrity and continuity of the extracted roads.

Description

Technical Field

[0001] This invention provides a remote sensing image road segmentation method that combines road detail feature extraction with context information perception, belonging to the field of remote sensing image road segmentation technology. Background Technology

[0002] Semantic segmentation is a fundamental task in computer vision, aiming to assign a semantic label category to each pixel in an image. It plays a crucial role in fields such as autonomous driving, medical imaging, and remote sensing. With societal development, extracting road information from remote sensing image data has gradually become a research hotspot. Road information can be used in many aspects of social life, such as urban planning and geographic information updates.

[0003] With the diversification of data sources and the increasing intelligence of algorithms, deep learning-based methods for road extraction from remote sensing images have become mainstream. Most deep learning methods have improved the ability to recover occluded roads in remote sensing images to varying degrees, but they still lose some image details and ignore the inherent properties of roads. This is because most studies, when improving skip connections, failed to design extraction modules specifically for road features, resulting in insufficient flexibility. Furthermore, these methods do not employ efficient attention mechanisms in context extraction, leading to limitations in handling background images. Considering these problems and causes, this invention proposes a remote sensing image road segmentation method that combines road detail feature extraction with context information awareness, improving road extraction capabilities while increasing the continuity of extracted roads. Summary of the Invention

[0004] To address the problems of existing road segmentation networks being susceptible to interference from background, occlusion, and similar road information, and struggling to extract complete and continuous roads, a remote sensing image road segmentation method combining road detail feature extraction and contextual information perception is proposed.

[0005] The technical solution adopted in this invention is: a remote sensing image road segmentation method combining road detail feature extraction and context information perception, comprising the following steps:

[0006] Step 1: Constructing the dataset: High-resolution remote sensing images are acquired through remote sensing satellites and divided into training and test sets according to the proportions. All images in the training set are preprocessed and data augmented. Both the training and test sets contain original remote sensing images and professionally labeled remote sensing images.

[0007] Step 2: Construct a road segmentation network model for remote sensing images: including:

[0008] Step 2.1: Construction of the feature encoding module;

[0009] Step 2.2: Construction of the road detail feature extraction module: The road detail feature extraction module includes a horizontal detail feature extraction module and a vertical detail feature extraction module for fusing multi-level strip features of the road. Both the horizontal and vertical detail feature extraction modules include strip spatial attention and effective channel attention with multiple scales for spatial and channel dimension selection.

[0010] Step 2.3: Construction of the context information awareness module: The context information awareness module is designed with a dilated convolution leader in a selective scanning block, and the dilation rate of the dilated convolution in the selective scanning block is changed to achieve the extraction of multi-scale detailed features. It is used for local multi-scale context extraction and global context extraction, and combined with effective channel attention for further feature filtering and fusion.

[0011] Step 2.4: Construction of the feature decoding module;

[0012] The feature encoding module and the feature decoding module achieve step-by-step feature fusion through the road detail feature extraction module; the fused features are then processed by the context information perception module to extract global and local multi-scale context features; the extraction results are used by the feature decoding module for decoding.

[0013] Step 3: Train the remote sensing image road segmentation network model;

[0014] Step 4: Result Prediction: Input the preprocessed remote sensing image into the trained remote sensing image road segmentation network model to obtain accurate road segmentation results.

[0015] Furthermore, the feature encoding module and the feature decoding module each consist of five stages. The feature encoding module includes, in sequence, a downsampling module, encoder 1, encoder 2, encoder 3 and encoder 4; the feature decoding module includes, in sequence, decoder 1, decoder 2, decoder 3, decoder 4 and a segmentation head.

[0016] Furthermore, the downsampling module consists of a 7×7 convolution operation with a stride of 2 and a 3×3 max pooling operation with a stride of 2; the four encoders are composed of stacked residual blocks of 3, 4, 6 and 3 respectively; each residual block is composed of two 3×3 convolution operations and residual concatenation.

[0017] Furthermore, each of the four decoders includes a 1×1 convolution that reduces the number of channels to 1 / 4, a 3×3 transposed convolution with a stride of 2, and a 1×1 convolution that doubles the number of channels. Each of the three convolutions is followed by batch normalization and the ReLU activation function. Finally, the output of the decoders is converted into road extraction results through a segmentation head, which consists of a 3×3 transposed convolution with a stride of 2, a 3×3 convolution, and a 2×2 convolution.

[0018] Furthermore, the multi-level features output by the encoder are fused step by step through three road detail feature extraction modules. Each road detail feature extraction module receives input from two parts: the output of the feature map at the current scale and the result of feature extraction at the previous scale. These two parts are processed separately and then weighted and added together as the input of the road detail feature extraction module.

[0019] Furthermore, the lateral detail feature extraction module includes a strip spatial attention consisting of two horizontal strip convolutional branches and one convolutional branch. The image input to the road detail feature extraction module is processed by horizontal strip convolutional kernels with kernels of 1×3 and 1×5, and then subjected to layer normalization. Simultaneously, the image input to the road detail feature extraction module is normalized by a 3×3 convolutional kernel layer, and then subjected to mean pooling and max pooling, respectively. The results are then summed after passing through an MLP layer, and the summed results are processed by Sigmoid to obtain a weighted value. The weighted value is then multiplied by the features obtained from the two horizontal strip convolutional branches, and then summed and passed through an output mapping layer and a series of feedforward networks to obtain the final horizontal output, wherein the output mapping layer adopts effective channel attention.

[0020] Furthermore, the vertical detail feature extraction module includes two vertical strip convolutional branches and one convolutional branch. The image input to the road detail feature extraction module is processed by vertical strip convolutional kernels with kernels of 3×1 and 5×1, and then normalized. Simultaneously, the image input to the road detail feature extraction module is normalized by a 3×3 convolutional kernel layer, and then subjected to mean pooling and max pooling, respectively. The results are then summed after passing through an MLP layer, and the summed results are processed by Sigmoid to obtain a weighted value. The weighted value is then multiplied by the features obtained from the two horizontal strip convolutional branches, and then summed and passed through an output mapping layer and a series of feedforward networks to obtain the final vertical output, wherein the output mapping layer uses effective channel attention.

[0021] Furthermore, the input of the context information perception module is the weighted sum of the output of encoder 4 and the output of the deepest road detail feature extraction module. The structure of the context information perception module is as follows: it consists of a location embedding block, four sets of dilated convolutions with dilation rates of 1, 3, 5 and 7, four 2D selective scanning blocks and effective channel attention. Each set of dilated convolutions and 2D selective scanning blocks is followed by element-wise multiplication with shared weights. Then the four sets of results are concatenated and output through effective channel attention.

[0022] Furthermore, the structure of the effective channel attention is as follows: it consists of a 3×3 convolution, a GeLU activation function, a layer normalization, and an efficient channel attention ECA.

[0023] Furthermore, the training process of the model in step three is as follows:

[0024] The preprocessed training set data is input into the remote sensing image road segmentation network model. The feature encoding module is initialized using the ResNet-34 weights pre-trained on the ImageNet-1K dataset, while the remaining parameters are randomly initialized. The remote sensing image road segmentation network model is trained until the model converges, and the model parameters are saved after training.

[0025] The advantages of this invention over the prior art are as follows:

[0026] (1) This invention employs residual blocks from the ResNet-34 network in the encoder section and initializes it using weights pre-trained on the ImageNet-1K dataset. This strategy not only effectively utilizes existing knowledge and significantly improves the encoder's performance, but also lays a solid foundation for subsequent remote sensing image road segmentation tasks. This method enables the model to possess a certain local feature extraction capability in the early stages of training, thereby accelerating the model's convergence speed and improving training efficiency.

[0027] (2) The present invention designs a road detail feature extraction module. This module integrates multi-level features through horizontal and vertical extraction, which alleviates the problem of edge detail information loss caused by continuous downsampling, reduces the semantic gap between the encoder and decoder, and makes the extracted road more continuous.

[0028] (3) The present invention designs a context information perception module, which effectively distinguishes between road and background information by dynamically extracting and integrating global and local multi-scale context information, making the extracted road structure more complete and continuous.

[0029] (4) The present invention designs a strip-shaped detail feature extraction method, which includes two parts: horizontal and vertical feature extraction and channel selection; horizontal and vertical feature extraction can extract the detailed features of the road at the current stage; channel selection selects the channel dimension of the feature map obtained after horizontal and vertical feature extraction.

[0030] (5) This invention designs a long-distance spatial selection method, which includes a location embedding block, four sets of dilated convolutions with dilation rates of 1, 3, 5 and 7, four 2D selective scanning blocks and channel attention. Each set of dilated convolutions and 2D selective scanning blocks is followed by element-wise multiplication with shared weights, and then the four sets of results are concatenated to obtain the output, thereby realizing spatial selection.

[0031] (6) In the feature decoding module, this invention upsamples the abstract feature map of the context information perception module to the original image resolution by transposing convolution. Specifically, this process not only effectively restores the details and spatial resolution of the feature map, but also combines the output of the detail feature extraction module, significantly enhancing the preservation of road spatial details and context information, and improving the model's ability to restore road details. Attached Figure Description

[0032] The present invention will be further described below with reference to the accompanying drawings:

[0033] Figure 1 This is a flowchart illustrating the remote sensing image road segmentation method proposed in this invention;

[0034] Figure 2 This is a schematic diagram of the structure of the remote sensing image road segmentation network model proposed in this invention;

[0035] Figure 3 This is a schematic diagram of the vertical detail feature extraction module proposed in this invention;

[0036] Figure 4 This is a schematic diagram of the lateral detail feature extraction module proposed in this invention;

[0037] Figure 5 This is a schematic diagram of the context information sensing structure proposed in this invention;

[0038] Figure 6 This is a schematic diagram of the effective channel attention structure proposed in this invention;

[0039] Figure 7 This is an example image of a remote sensing image from the DeepGlobe dataset used in this embodiment of the invention;

[0040] Figure 8 This is an example image of a remote sensing image with labels from the DeepGlobe dataset used in this embodiment of the invention;

[0041] Figure 9 It was extracted using the method of this invention. Figure 6 A schematic diagram of the road structure in China. Detailed Implementation

[0042] like Figures 1 to 9 As shown, this invention provides a remote sensing image road segmentation method that combines road detail feature extraction and context information awareness, including the following steps:

[0043] Step 1: Dataset Construction: High-resolution remote sensing images of 1024×1024 pixels were acquired via remote sensing satellites and divided into training and test sets at an 8:2 ratio. This dataset includes the original remote sensing images and professionally labeled data. Subsequently, a series of preprocessing and data augmentation operations were performed on the images and labels in the training set, including horizontal, vertical, and joint flipping, rotation, scaling, random noise injection, and color adjustment. These augmentation methods increased the diversity of image perspectives, improved the model's robustness to road structures under different orientations and conditions, and thus enhanced the accuracy and reliability of the remote sensing image road segmentation network. The processed images were used to train the network model to achieve accurate segmentation of road structures in remote sensing images, laying the foundation for subsequent analysis.

[0044] Step 2: Model Construction: The remote sensing image road segmentation network model includes a feature encoding module (Encoder), a road details feature extraction module (RDFEM), a context-aware fusion module (CAFM), and a feature decoding module (Decoder). The feature encoding and decoding modules each consist of five stages. The feature encoding module includes a downsampling module, encoder 1, encoder 2, encoder 3, and encoder 4; the feature decoding module includes decoder 1, decoder 2, decoder 3, decoder 4, and a segmentation head. The feature encoding and decoding modules achieve progressive feature fusion through the road details feature extraction module. The fused features are then processed by the context-aware fusion module to extract global and local multi-scale contextual features. The extracted results are used by the feature decoding module for decoding.

[0045] The specific steps for building a road segmentation network model from remote sensing images are as follows:

[0046] Step 2.1: Construction of the feature encoding module: A ResNet-34 network pre-trained on the ImageNet-1K dataset was used to construct a downsampling module, encoder 1, encoder 2, encoder 3 and encoder 4 for feature extraction, resulting in feature maps of four stages. These maps will then be used as input to the bidirectional multi-level road feature dynamic fusion module for bidirectional dynamic fusion.

[0047] The structure of the feature encoding module is as follows:

[0048] The feature encoding module comprises a downsampling module, encoder 1, encoder 2, encoder 3, and encoder 4. The downsampling module consists of a 7×7 convolution operation with a stride of 2 and a 3×3 max pooling operation with a stride of 2. The four encoders are composed of stacked residual blocks of 3, 4, 6, and 3, respectively. Each residual block consists of two 3×3 convolution operations and residual concatenation. The four encoders generate four feature maps, each reflecting the feature representation of the image at different levels of abstraction. Specifically, the feature map generated by encoder 1 is a shallow feature map, which mainly captures basic and local details in the image; while the feature map generated by encoder 4 is a deep feature map, which contains more abstract and global information, and is particularly crucial for performing fine-grained semantic segmentation tasks.

[0049] Step 2.2: Construction of the Road Detail Feature Extraction Module: As the encoder continuously downsamples, some image details are gradually lost. Therefore, it is crucial to extract and save the features displayed by each road layer in a timely manner. Skip connections play a vital role, serving as the carrier between the encoder and decoder. Shallow features contain rich road shape and texture features, which are important for edge feature recovery. Meanwhile, the road's topological structure means that simple feature extraction cannot effectively supplement and improve model performance. Only by designing a unique and efficient feature extraction module tailored to the data characteristics can the road shape be recovered to a great extent.

[0050] The structure of the road detail feature extraction module is as follows:

[0051] Three road detail feature extraction modules are used to progressively fuse the multi-level features output by the encoder. Each module receives input from two parts: the output feature map at the current scale and the result of feature extraction from the previous scale. These two parts are processed separately and then weighted and summed as the input to this module. The input goes through two parallel parts, which extract features from the horizontal and vertical directions respectively.

[0052] Figure 3 This diagram illustrates the extraction of vertical features. The input then undergoes three parts: the image is processed by vertical strip convolutional kernels of 3×1 and 5×1, followed by layer normalization to obtain V1 and V2, respectively. After normalization by a 3×3 convolutional kernel layer, the input undergoes average pooling (AP) and max pooling (MP), respectively. The results are then summed after passing through an MLP layer, and a sigmoid function is applied to the sum to obtain a weighted value. This weighted value is then multiplied by the features obtained from the two branches, and the sums are passed through an output mapping layer (effective channel attention) and a series of feedforward networks to obtain the final vertical output.

[0053] Figure 4 This diagram illustrates the extraction of horizontal features. The input then goes through three parts: the image is processed by horizontal strip convolutional kernels of 1×3 and 1×5, followed by layer normalization to obtain H1 and H2, respectively. After normalization by a 3×3 convolutional kernel layer, the input undergoes average pooling (AP) and max pooling (MP), respectively. The results are then summed after passing through an MLP layer, and the sum is processed by a sigmoid function to obtain a weighted value. This weighted value is then multiplied by the features obtained from the two branches, summed, and then passed through the output mapping layer Projection (i.e., effective channel attention, such as...). Figure 6 (As shown) and a series of feedforward networks to obtain the final horizontal output.

[0054] This module integrates information from two scales as input, which can retain the road information extracted from the previous scale. At the same time, the output of this module not only integrates the road strip features in two directions, but also provides detailed road information from multiple scales using strip convolution kernels of different sizes in each direction. This can effectively supplement the decoder to provide comprehensive and accurate road structure features.

[0055] Step 2.3: Construction of the Context Information Awareness Module: In remote sensing image road segmentation tasks, a large number of road pixels are obscured by surrounding buildings, trees, and shadows. Extracting only road features cannot guarantee the completeness of the extracted road. Certain objects in the background have connections with the road and can, to some extent, guide road extraction. Capturing the relationship between the road and its surrounding environment between the encoder and decoder can improve the overall representation of the extracted road, thereby increasing the accuracy of prediction results. This invention combines the advantages of dilated convolution and 2D selective scanning to design a dilated convolution-led selective scanning block for dynamically extracting and integrating local multi-scale and global contextual information. This effectively distinguishes between road and background information, making the extracted road structure more complete and continuous. Then, effective channel attention filtering is used to filter out road-like interference information and suppress background and other interference information, ultimately constructing a dual-context dynamic extraction module.

[0056] The structure of the context information awareness module is as follows:

[0057] To compensate for information loss caused by stepwise downsampling, this invention employs a strategy of weighted summation of the output of encoder 4 and the output of the deepest road detail feature extraction module as the input to this module. The input is doubled in number of channels by a mapping layer and then divided into two parts in the channel dimension, x and z. x is used for subsequent feature extraction and long sequence modeling, while z is multiplied by the long sequence modeling result as shared weights after passing through the SiLU function. Dilated convolution combined with positional encoding, layer normalization, and feedforward network layers, and embedded with 2D selective scanning as a spatial attention mechanism, constructs a more robust block (dilated convolution-leading selective scanning block). It is worth noting that, due to the need to remember long sequence dependencies, the number of hidden states in the state space equation is usually large, which may lead to channel redundancy, thus hindering the learning of key channel representations. Therefore, this invention implements channel attention in the mapping layer of the dilated convolution-leading selective scanning block to solve this problem. At the same time, by changing the dilation rate of the dilated convolution in the dilated convolution-leading selective scanning block, multi-scale detail feature extraction can be achieved. This invention employs progressively dilated convolutions (dilation rates of 1, 3, 5, and 7) to extract multi-scale road detail features. Subsequently, 2D selective scanning is used to capture detailed features and long-range contextual dependencies of pixels surrounding the road. These four parallel dilated convolution-leading selective scanning blocks are performed in one-quarter of the original input channels. Finally, the outputs of the four channels are concatenated and passed through a final output mapping layer (effective channel attention) to obtain the module's output, generating a fused feature map with rich detail features and long-range contextual information. Simultaneously, the final output of the context information awareness module serves as input to the decoder, promoting more refined prediction reconstruction and ensuring the accuracy and consistency of the segmentation results.

[0058] Step 2.4: Construction of the Feature Decoding Module: The feature decoding module aims to convert the abstract feature maps output by the dual-context dynamic extraction module back to the same spatial resolution as the original image, thereby achieving accurate semantic segmentation of the remote sensing image. In this step, it starts with the global-local context extraction features aggregated by the dual-context dynamic extraction module, performs upsampling through transposed convolution, and combines the output from the corresponding stage of the bidirectional multi-level road feature dynamic fusion module to supplement detailed boundary information and contextual information of the road, thereby improving the accuracy and precision of semantic segmentation.

[0059] The structure of the feature decoding module is as follows:

[0060] Each of the four decoders includes a 1×1 convolution that reduces the number of channels to 1 / 4, a 3×3 transposed convolution with a stride of 2, and a 1×1 convolution that doubles the number of channels. Each of the three convolutions is followed by batch normalization and the ReLU activation function. Finally, the decoder output is converted into road extraction results by a segmentation head, which consists of a 3×3 transposed convolution with a stride of 2, a 3×3 convolution, and a 2×2 convolution.

[0061] Step 3: Model Training: Input the preprocessed training set data into the remote sensing image road segmentation network; initialize the feature encoding module with the weights of the ResNet-34 model pre-trained on the ImageNet-1K dataset, and randomly initialize the remaining network parameters; train the remote sensing image road segmentation network model until the model converges; after training, save the trained remote sensing image road segmentation network model parameters.

[0062] Step 4: Result Prediction: Input the preprocessed remote sensing road image into the trained remote sensing image road segmentation network model to obtain accurate segmentation results of the remote sensing image data.

[0063] The training program is built on the PyTorch framework, using the Adam algorithm as the optimizer. The loss function combines binary cross-entropy and the Dice coefficient to alleviate class imbalance and enhance the model's segmentation ability for small-scale targets. The initial learning rate is set to 2e-4, the batch size to 4, and the maximum training batch size to 500. The learning rate adjustment mechanism uses the Plateau Learning Rate Scheduler, with a decay factor of 0.5 and a tolerance period of 5, enabling dynamic adjustment of the learning rate based on the learning process to avoid local optima. Furthermore, an early stopping condition is implemented: if the loss value does not decrease significantly for eight consecutive epochs, the model is considered to have converged, and the training process is terminated early. This aims to promptly curb overfitting while reasonably controlling computational costs.

[0064] The formula for calculating the binary cross-entropy loss function is as follows:

[0065]

[0066] In the above formula: y is the actual pixel label value, y' is the predicted label pixel value, and N is the number of label categories;

[0067] The formula for calculating the Dice coefficient loss function is:

[0068]

[0069] In the above formula: X is the generated prediction graph, Y is the true label, |X∩Y| is the intersection between the label and the prediction, |X| is the number of elements in the label, and |Y| is the number of elements in the prediction;

[0070] The final semantic segmentation loss function is the weighted sum of the cross-entropy loss function and the Dice coefficient loss function, calculated as follows:

[0071] L s =L dice +L cross .

[0072] The present invention will be further described below with reference to specific embodiments.

[0073] The experimental dataset used was the DeepGlobe Road Extraction Dataset, which comes from the Remote Sensing Road Extraction Challenge held at CVPR 2018. The dataset contains 6226 sets of remote sensing images (1024×1024 pixels) with road labels, at a resolution of 0.5 m / pixel. The images were collected by Digital Globe satellites and cover a wide geographical area, including cities and rural areas in multiple countries, as well as various types of road and street networks. Furthermore, to maintain the data's authenticity and practicality, the authors deliberately omitted labels for paths in farmland, preventing the model from making incorrect extractions during recognition. This makes the dataset more challenging and practical, and provides a broader scope for exploration.

[0074] In evaluating the road segmentation performance of the model, this invention uses four commonly used evaluation metrics in semantic segmentation, and the evaluation results are shown in Table 1. The four evaluation metrics are precision, recall, F1 score, and intersection-over-union ratio (IoU). These metrics can comprehensively and objectively reflect the model's performance on the road segmentation task. Through analysis of... Figure 6 (Original remote sensing road image) Figure 7 (Remote sensing road segmentation reference image with manual annotation) and Figure 8 A comparative analysis of the predicted road segmentation results (images obtained after applying the method of this invention) shows that the method of this invention demonstrates excellent performance in road extraction from remote sensing images. Specifically, Figure 8 The predicted segmentation results shown are consistent with Figure 7 The high degree of consistency among the manually labeled standard images strongly demonstrates the accuracy and reliability of the method of the present invention in road segmentation tasks.

[0075] accuracy Recall rate F1 score Intersection and Union 84.58％ 84.59％ 84.58％ 73.29％

[0076] Table 1 shows the specific metrics on the DeepGlobe road extraction dataset.

[0077] This invention discloses an innovative road segmentation method for remote sensing images, which integrates road detail feature extraction and contextual information awareness. The invention aims to achieve highly accurate segmentation of road regions in high-resolution remote sensing images through advanced deep learning technology combined with a specially designed network architecture and modules. First, detailed manually labeled images are created based on high-resolution remote sensing images with RGB three channels, serving as a standard reference for model training. Next, the complete image dataset is proportionally divided into training and test sets, and a series of preprocessing operations, such as denoising and feature enhancement, are performed on the images in the training set to optimize image quality and lay a solid foundation for model training. In the model building stage, this invention creatively designs a road segmentation method for remote sensing images that combines road detail feature extraction and contextual information awareness. Specifically, the road detail feature extraction module aggregates the features of the road itself from both horizontal and vertical directions, passing them as supplementary details to the decoder. This effectively compensates for the loss of spatial information and edge details caused by continuous downsampling, enhancing the model's understanding of complex scenes and multi-layered road structures. On the other hand, the context information awareness module achieves the fusion of long-distance and local multi-scale context information, which helps to suppress background interference, occlusion effects, and confusion caused by similar road structures, ensuring the integrity and continuity of the extracted road structure. During model training, this invention inputs the preprocessed training images and their corresponding label images into the road segmentation model, and uses optimization algorithms to continuously adjust the network parameters until the model can accurately identify and segment road regions. When the model reaches convergence, the optimal parameter settings are saved for subsequent applications. Finally, this invention can automatically generate predicted label images, i.e., segmented road images, by inputting new remote sensing images into the trained model. Compared with existing technologies, the unique feature of this invention lies in its application of road detail feature extraction and context information awareness mechanisms, which not only improves the connectivity and accuracy of road segmentation, but also enhances the model's ability to distinguish different levels of roads and recover edge details.

[0078] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A remote sensing image road segmentation method combining road detail feature extraction and contextual information perception, characterized in that: Includes the following steps: Step 1: Constructing the dataset: High-resolution remote sensing images are acquired through remote sensing satellites and divided into training and test sets according to the proportions. All images in the training set are preprocessed and data augmented. Both the training and test sets contain original remote sensing images and professionally labeled remote sensing images. Step 2: Construct a road segmentation network model from remote sensing images: including: Step 2.1: Construction of the feature encoding module; Step 2.2: Construction of the road detail feature extraction module: The road detail feature extraction module includes a horizontal detail feature extraction module and a vertical detail feature extraction module for fusing multi-level strip features of the road. Both the horizontal and vertical detail feature extraction modules include strip spatial attention and effective channel attention with multiple scales for spatial and channel dimension selection. The horizontal detail feature extraction module includes a strip spatial attention consisting of two horizontal strip convolutional branches and one convolutional branch. The image input to the road detail feature extraction module is processed by horizontal strip convolutional kernels with kernels of 1×3 and 1×5, and then normalized. Simultaneously, the image input to the road detail feature extraction module is normalized by a 3×3 convolutional kernel layer, and then subjected to mean pooling and max pooling, respectively. The results are then summed after passing through an MLP layer, and the sum is processed by Sigmoid to obtain a weighted value. The weighted value is then multiplied by the features obtained from the two horizontal strip convolutional branches, and then summed. The result is then passed through an output mapping layer and a series of feedforward networks to obtain the final horizontal output, where the output mapping layer uses effective channel attention. Step 2.3: Construction of the context information awareness module: The context information awareness module is designed with a dilated convolution leader in a selective scanning block, and the dilation rate of the dilated convolution in the selective scanning block is changed to achieve the extraction of multi-scale detailed features, which is used for local multi-scale context extraction and global context extraction, and combined with effective channel attention for further feature filtering and fusion. Step 2.4: Construction of the feature decoding module; The feature encoding module and the feature decoding module achieve step-by-step feature fusion through the road detail feature extraction module; the fused features are then processed by the context information perception module to extract global and local multi-scale context features; the extraction results are used by the feature decoding module for decoding. Step 3: Train the remote sensing image road segmentation network model; Step 4: Result Prediction: Input the preprocessed remote sensing image into the trained remote sensing image road segmentation network model to obtain accurate road segmentation results.

2. The remote sensing image road segmentation method combining road detail feature extraction and context information perception according to claim 1, characterized in that: The feature encoding module and the feature decoding module each consist of five stages. The feature encoding module includes, in sequence, a downsampling module, encoder 1, encoder 2, encoder 3 and encoder 4; the feature decoding module includes, in sequence, decoder 1, decoder 2, decoder 3, decoder 4 and a segmentation head.

3. The remote sensing image road segmentation method combining road detail feature extraction and context information perception according to claim 2, characterized in that: The downsampling module consists of a 7×7 convolution operation with a stride of 2 and a 3×3 max pooling operation with a stride of 2; the four encoders are composed of stacked residual blocks of 3, 4, 6 and 3 respectively; each residual block is composed of two 3×3 convolution operations and residual concatenation.

4. The remote sensing image road segmentation method combining road detail feature extraction and context information perception according to claim 2, characterized in that: Each of the four decoders includes a 1×1 convolution that reduces the number of channels to 1 / 4, a 3×3 transposed convolution with a stride of 2, and a 1×1 convolution that doubles the number of channels. Each of the three convolutions is followed by batch normalization and the ReLU activation function. Finally, the output of the decoders is converted into road extraction results by a segmentation head, which consists of a 3×3 transposed convolution with a stride of 2, a 3×3 convolution, and a 2×2 convolution.

5. The remote sensing image road segmentation method combining road detail feature extraction and context information perception according to claim 2, characterized in that: Three road detail feature extraction modules are used to perform step-by-step fusion of multi-level features output by the encoder. Each road detail feature extraction module receives input from two parts: the output of the feature map at the current scale and the result of feature extraction at the previous scale. These two parts are processed separately and then weighted and added together as the input of the road detail feature extraction module.

6. The remote sensing image road segmentation method combining road detail feature extraction and context information perception according to claim 5, characterized in that: The vertical detail feature extraction module includes two vertical strip convolutional branches and one convolutional branch. The image input to the road detail feature extraction module is processed by vertical strip convolutional kernels with kernels of 3×1 and 5×1, and then normalized. Simultaneously, the image input to the road detail feature extraction module is normalized by a 3×3 convolutional kernel layer, and then subjected to mean pooling and max pooling respectively. The results are summed after passing through an MLP layer, and then the summed result is processed by Sigmoid to obtain a weighted value. The weighted value is then multiplied by the features obtained from the two vertical strip convolutional branches, and then summed. The summed result is then passed through an output mapping layer and a series of feedforward networks to obtain the final vertical output, where the output mapping layer uses effective channel attention.

7. The remote sensing image road segmentation method combining road detail feature extraction and context information perception according to claim 2, characterized in that: The input to the context information perception module is the weighted sum of the output of encoder 4 and the output of the deepest road detail feature extraction module. The structure of the context information perception module is as follows: it consists of a location embedding block, four sets of dilated convolutions with dilation rates of 1, 3, 5 and 7, four 2D selective scanning blocks and effective channel attention. Each set of dilated convolutions and 2D selective scanning blocks is followed by element-wise multiplication with shared weights. Then the four sets of results are concatenated and output through effective channel attention.

8. The remote sensing image road segmentation method combining road detail feature extraction and context information perception according to claim 7, characterized in that: The structure of the effective channel attention is as follows: it consists of a 3×3 convolution, a GeLU activation function, a layer normalization, and an efficient channel attention (ECA).

9. The remote sensing image road segmentation method combining road detail feature extraction and context information perception according to claim 1, characterized in that: The training process of the model in step three is as follows: The preprocessed training set data is input into the remote sensing image road segmentation network model. The feature encoding module is initialized using the ResNet-34 weights pre-trained on the ImageNet-1K dataset, while the remaining parameters are randomly initialized. The remote sensing image road segmentation network model is trained until the model converges, and the model parameters are saved after training.

Images