An occlusion-aware road intelligent extraction method

Abstract

This invention discloses an intelligent road extraction method based on occlusion perception. The method first acquires remote sensing images of the road area to be extracted. Through multi-level depth analysis, a feature map integrating global and local information is obtained. The encoded feature map is then used with horizontal, vertical, and double diagonal strip convolution operations to establish the correlation between intersecting channels. Combining multi-directional features and intersecting channel correlations, attention weights are generated to obtain optimized features. Finally, a decoder produces deep features, which are fused with the optimized features using skip connections. The method then assesses the rationality of the morphological structure perception constraint prediction results obtained by combining segmentation base loss, structural continuity loss, and boundary accuracy loss, and completes automatic road extraction based on multi-level supervision. This invention is applicable to automatic road information extraction in fields such as autonomous driving, urban planning, and geographic information system construction.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing data processing technology, and in particular relates to an intelligent road extraction method based on occlusion perception. Background Technology

[0002] As a core connecting link between multifunctional urban areas, road networks play an irreplaceable role in residents' daily activities and socio-economic development. Their complex spatial structure and diverse morphological characteristics make traditional manual on-site surveying and marking methods insufficient in terms of efficiency and accuracy, especially revealing significant technological limitations in large-scale road network construction. In recent years, automatic road extraction technology, utilizing high-resolution remote sensing imagery data and integrating machine learning and computer vision algorithms, has gradually become a research hotspot in the international academic community. This technology not only significantly improves the efficiency of route planning but also plays a crucial role in areas such as autonomous driving, smart city planning, geographic information system maintenance, and emergency response.

[0003] However, with the continuous development of remote sensing technology, image data has become more diverse and complex. Some narrow roads have blurred boundaries due to obstruction by surrounding buildings or vegetation, and the road surface material characteristics are mixed with the types of surrounding landforms (construction sites, farmland, etc.). Furthermore, traditional methods suffer from inefficiency and insufficient multi-scale fusion capabilities in practice, failing to comprehensively capture the multi-scale features of roads. In this context, improving the intelligence level of path analysis, expanding its applicability, and enhancing accuracy become extremely urgent and important. Summary of the Invention

[0004] To address the aforementioned problems, this invention proposes an intelligent road extraction method based on occlusion perception. This method aims to solve the problem that traditional methods cannot fully capture the multi-scale features of roads in practical operation, and are difficult to extract roads in situations with blurred boundaries and mixed ground features. This greatly improves the accuracy and robustness of road recognition in complex scenarios.

[0005] The above objectives are achieved through the following technical solutions:

[0006] A method for intelligent road extraction based on occlusion perception, comprising the following steps:

[0007] S1. Dense coding feature extraction and multi-scale fusion: Obtain remote sensing images of the road area to be extracted. Based on the Link-Net backbone network architecture, combine residual coding and dense convolutional units (DCUs) to directly transfer shallow features to deep layers, force feature reuse, reduce information loss and enhance expressive power. At the end of the network, a multi-scale fusion module is introduced to obtain a feature map that integrates global and local information through multi-level deep analysis.

[0008] S2. Double cross-shaped attention optimization: The feature map obtained by encoding in step S1 is used to establish the correlation between the cross channels by using strip convolution operations in the horizontal, vertical and double diagonal directions. Combining multi-directional features and cross channel correlation, attention weights are generated to obtain optimized features.

[0009] S3. Multi-level supervised prediction for morphological structure perception: The deep features generated by the decoder are fused with the optimized features by skip connections. The rationality of the morphological structure perception constraint prediction results obtained by combining the segmentation base loss, structural continuity loss and boundary accuracy loss is evaluated, and the automatic extraction of roads is completed based on multi-level supervision.

[0010] Furthermore, the dense coding feature extraction and multi-scale fusion described in step S1 specifically includes the following sub-steps:

[0011] S11. Acquire remote sensing images of the road area to be extracted;

[0012] S12. Based on the Link-Net backbone network architecture, residual coding and dense feature enhancement are combined to directly pass shallow features to deep layers, forcing feature reuse. Specifically, dense convolutional units (DCUs) are embedded in the concatenation of residual coding units, forming an alternating structure of "residual coding unit → dense convolutional unit (DCU) → residual coding unit". First, the input feature F is processed by one residual coding unit. in Perform a preliminary transformation and output F. resi Then, F resi As the input to the Dense Convolutional Unit (DCU), the input to each subsequent DCU is the concatenation of the output of the previous DCU and the output of the previous residual unit. Therefore:

[0013] DCU u+1 =Concat(F resi DCU i )

[0014] Among them, DCU i+1 This represents the output of the (i+1)th dense convolutional unit (DCU), ultimately yielding multi-resolution, multi-semantic hierarchical features.

[0015] The densely enhanced features are compressed in dimension by a 1×1 convolution and then input into the next residual encoding unit, and so on.

[0016] S13. The multi-scale information fusion module is located at the end of the network and is responsible for the deep integration of multi-resolution and multi-semantic level features output from the encoding stage. Specifically:

[0017] S131. The multi-resolution and multi-semantic hierarchical features output during the encoding stage are processed using different convolution kernels. Local details are captured by 3×3 convolution, continuous features of road segments are captured by 5×5 convolution, and global topological features are captured by 7×7 convolution. The convolution results are then summed and fused to integrate features at different scales.

[0018] S132. Aggregate the fused features of different scales obtained in S131 using average pooling and max pooling, and obtain local attention through 7×7 convolution and sigmoid activation function;

[0019] S133. The spatial dimension of the multi-resolution, multi-semantic hierarchical features output in the encoding stage is compressed to 1×1, global average pooling is performed to extract global information; and attention maps are generated using 1×1 convolution, batch normalization and ReLU activation function, and then their size is expanded to match the dimension of the input feature map to obtain global attention.

[0020] S134. Integrate the results of global attention and local attention to obtain the final output feature map.

[0021] Furthermore, the double-cross bar attention optimization described in step S2 specifically includes the following sub-steps:

[0022] S21. Feature Extraction via Multi-directional Strip Convolution: The feature map output from step S1 is processed using four types of strip convolution kernels: horizontal, vertical, 45° diagonal, and 135° diagonal, to capture road features in different directions. The horizontal strip convolution kernel H has a size of 1×k, where k represents the receptive field size, ranging from [3,15], and slides only in the horizontal direction, responding to horizontally extending roads. The vertical strip convolution kernel V has a size of k×1, responding to vertically extending roads. The 45° diagonal strip convolution kernel D1 has a size of k×k, with weights concentrated in the 45° diagonal direction, responding to diagonal roads. The 135° diagonal strip convolution kernel D2 has a size of k×k, with weights concentrated in the 135° diagonal direction, responding to diagonally opposite roads.

[0023] For the input feature map C represents the number of channels, and H and W represent the height and width, respectively. After four types of strip convolution, feature maps in four directions are obtained:

[0024] F H =Conv H (F),F V =Conv V (F),F D1 =Conv D1 (F),F D2 =Conv D2 (F)

[0025] Among them, F H Conv represents the horizontal feature map output after horizontal strip convolution. H (F) represents a horizontal convolution operation on the feature map F, where F V Conv represents the feature map in the vertical direction output after vertical strip convolution. V (F) denotes a vertical convolution operation on the feature map F, where F D1 Conv represents the feature map output along the 45° diagonal direction after a linear convolution with a 45° diagonal line. D1 (F) indicates that a 45° diagonal convolution operation is performed on the feature map F, where F D2 Conv represents the feature map output along the 135° diagonal direction after a linear convolution along the 135° diagonal. 22 (F) indicates that a 135° diagonal convolution operation is performed on the feature map F;

[0026] S22. Channel correlation modeling of cross-correlation: The horizontal feature map F output after performing horizontal strip convolution on step S21. H By compressing the spatial dimension into a vector, we obtain the channel feature vector. Where c is the channel index, the cross-correlation between any two channels c and c′ is:

[0027]

[0028] Among them, Corr H (c,c′) represents the cross correlation between any two channels c and c′ in the feature map in the horizontal direction; This represents the feature vector of channel c; represents the feature vector of channel c′; ‖·‖ represents the norm; · is the inner product, and the result ranges from [-1,1]. The larger the value, the higher the consistency of the features of the two channels in the horizontal direction.

[0029] Similarly, for the feature map F obtained in the vertical direction in step S21 V Feature diagram F along the 45° diagonal direction D1 Feature diagram F along the 135° diagonal direction D2 The spatial dimension is compressed into vectors, and the cross-correlation of any two channels c and c′ is calculated to obtain the cross-correlation Corr of any two channels c and c′ in the vertical feature map. V The cross-correlation between any two channels c and c′ in the feature map along the 45° diagonal direction (c, c′) Corr D1 The cross-correlation between any two channels c and c′ in the feature map along the 135° diagonal direction (c, c′) Corr D2(c,c′), taking the average of the correlation matrices in these four directions, yields the comprehensive channel correlation matrix M(c,c′):

[0030]

[0031] The larger M(c,c′) is, the stronger the feature consistency of channels c and c′ in the four directions, and they may jointly encode road features.

[0032] S23. Based on the comprehensive channel correlation matrix M(c,c′) obtained in step S22, calculate the importance of each channel and obtain channel attention:

[0033]

[0034] Where CA(c) represents the attention for channel c, M(c,c′) represents the correlation matrix between channels c and c′, σ is the Sigmoid function, and W c v is a learnable parameter c CS(c) is the channel vector of the original feature map. The higher the CS(c) value, the more significant the road features contained in the channel.

[0035] The feature maps from the four directions are weighted and fused to generate a spatial attention map:

[0036] SA(x,y)=W(W s ·(W H (x,y)+F V (x,y)+F D1 (x,y)+F D2 (x,y)))

[0037] Where (x, y) are spatial coordinates, Ws s As a learnable parameter, σ outputs a weight of [0,1], where a higher value indicates a greater likelihood that the location belongs to a road.

[0038] The final attention map is achieved by multiplying channel attention and spatial attention: Attention(c,x,y)=CA(c)·SA(x,y).

[0039] Furthermore, the multi-level supervised prediction of morphological structure perception described in step S3, wherein morphological structure perception is obtained by combining structural continuity loss and boundary accuracy loss with the basic loss, the structural continuity loss quantifies the deviation between the predicted road and the real road in terms of skeleton shape, connectivity, and geometric dimensions, forcing the model to generate a continuous and unbroken road structure; the boundary accuracy loss strengthens the model's attention to road edge pixels, reduces the spatial deviation between the predicted boundary and the real boundary, and ensures that the boundary between the road and the background is clearly distinguishable; the basic loss L baseThe intersection-overlap ratio (IoU) measures the degree of overlap between the predicted and actual roads. Y represents the true label, P represents the predicted probability map, and the overall objective function L is as follows:

[0040] L=λ base L base +λ struct L struct +λ bound L bound

[0041] L base L struct and L bound These represent the basic loss, structural continuity loss, and boundary accuracy loss, respectively, which are combined to form a multi-objective constraint, λ. base , λ struct and λ boudn These are the weights corresponding to the basic loss, structural continuity loss, and boundary accuracy loss, respectively; the specific design methods for the basic loss, structural continuity loss, and boundary accuracy loss are as follows:

[0042] (1) Structural continuity loss design:

[0043] The structural continuity loss is a weighted combination of three aspects: skeleton consistency, regional connectivity, and geometric dimensions, addressing continuity constraints in different dimensions:

[0044] L struct =L sk +L cn +L sp

[0045] Among them, L sk L represents the skeleton consistency loss. cn L represents the regional connectivity loss. sp Indicates geometric loss;

[0046] The road skeleton, specifically the road's centerline, is the core carrier of continuity. By comparing the spatial distance between the predicted skeleton and the actual skeleton, it is ensured that the centerline does not shift or break. To this end, distance transformation is used to perform morphological refinement on both the actual label Y and the binarized predicted road probability map P, resulting in the actual skeleton S. Y and predicted skeleton S P Then, there is:

[0047]

[0048] Wherein d(p,S) P ) represents the Euclidean distance from pixel p to the nearest pixel in set S, and |S| represents the total number of skeleton pixels, which is used to penalize skeleton offset and breakage;

[0049] Road breaks lead to an increase in connected components. To reduce "fragmented" regions in the prediction results, the number and size of connected components are constrained. Therefore, connected component analysis is performed on the binarized predicted road probability map P, resulting in K connected domains C1, C2, ..., CK. K Its area is A1, A2, ..., A K Furthermore, the number of actual connected components K is obtained by analyzing the connected components of the real road Y. * and average area Then we have:

[0050]

[0051] Where λ1 and λ2 are weights and λ1>λ2, the number of fractures is constrained first, thereby penalizing excessive fractures or small fragments.

[0052] Geometric dimensions are used to constrain road width consistency and orientation smoothness. Width consistency is obtained by calculating the deviation between the predicted width at each point on the road and the actual width, i.e., the distance perpendicular to the skeleton. Orientation smoothness is measured by the change in the orientation angle of the skeleton points, penalizing sharp turns in non-intersection scenarios. Therefore:

[0053]

[0054] Among them, w P (j) and w Y (j) represents the predicted width and the actual road width at skeleton point j, respectively, and θ j (2) Design of boundary accuracy loss: (i) the tangent direction angle of skeleton point j;

[0055] Boundary accuracy loss focuses on pixel-level errors in the road edge region, combining boundary distance metrics and edge attention mechanisms:

[0056] L bound =L hd +K ed

[0057] Among them, K hd K represents the boundary distance loss. ed This indicates peripheral attention loss;

[0058] Boundary distance loss is applied to the ground truth label Y and the predicted probability map P using the Canny edge detection operator to obtain the ground truth boundary B. Y and prediction boundary B P And by using the average Hausdorff distance to directly quantify the degree of boundary offset, and forcing the predicted boundary to move closer to the true boundary, we have:

[0059]

[0060] Where d(x,B) is the shortest Euclidean distance from pixel p to the boundary set B.

[0061] Edge attention loss assigns higher weights to boundary pixels, enhancing the model's focus on edge regions (boundary pixels have a low proportion in the overall picture but a significant impact on visual integrity). This is achieved through the study of true boundaries (B). Y Generate a weighted graph W(x), and introduce this weight into the cross-entropy, then we have:

[0062]

[0063] Here, y(x) is the true label of pixel x, and p(x) is the predicted probability, in order to improve the classification accuracy of the boundary region and avoid edge pixels being "diluted" by background pixels.

[0064] The advantages of this invention compared to the prior art are:

[0065] 1. This invention effectively solves the problem of feature information loss by using dense coding feature extraction and multi-scale fusion, improves the ability to represent the spatial distribution details of the target, and thus better captures multi-scale contextual information and improves segmentation accuracy.

[0066] 2. This invention employs a double cross-shaped attention module to focus on the interaction characteristics of road pixels in the horizontal, vertical, and diagonal directions. It models multi-directional pixel dependencies from a global perspective, breaking through the limitations of traditional multi-scale integration methods in capturing long-distance dependencies and improving feature selection accuracy.

[0067] 3. The multi-level supervision of morphological structure perception in this invention integrates structural continuity and boundary accuracy loss to complete the automatic extraction of roads, and ensures that each stage of the network contributes to the final segmentation extraction result, promoting more refined and accurate output. Attached Figure Description

[0068] Figure 1 This is the overall process of the present invention;

[0069] Figure 2 This is the technical architecture of the present invention;

[0070] Figure 3 This is a structural diagram of the dense multi-scale fusion module of the present invention;

[0071] Figure 4 This is a schematic diagram of the double cross-shaped attention optimization module of the present invention;

[0072] Figure 5 This is the CHN6-CUG road dataset used in the embodiments of the present invention;

[0073] Figure 6This is a visual comparison of the method of the present invention with other methods. Figure 6 In the image, (a) represents four original images, (b) represents the ground truth of the four original images in (a), (c) represents the result extracted from the four original images in (a) using the Segmenter method, (d) represents the result extracted from the four original images in (a) using the UNetformer method, and (e) represents the result extracted from the four original images in (a) using the method of this invention.

[0074] Figure 7 To conduct comparative experiments between the attention module of this invention and other attention modules. Figure 7 In the image, (a) represents four original images, (b) represents the ground truth of the four original images in (a), (c) represents the result extracted from the four original images in (a) using the SE channel attention module, (d) represents the result extracted from the four original images in (a) using the DA dual attention module, (e) represents the result extracted from the four original images in (a) using the CBAM spatial channel attention module, (f) represents the result extracted from the four original images in (a) using the CCA channel cross attention module, (g) represents the result extracted from the four original images in (a) using the improved CCA module, and (h) represents the result extracted from the four original images in (a) using the double cross bar attention module of this invention. Detailed Implementation

[0075] like Figure 1 As shown, this invention provides an intelligent road extraction method based on occlusion perception, mainly comprising three steps: dense coding feature extraction and multi-scale fusion, double-cross bar attention optimization, and multi-level supervised prediction based on morphological structure perception. Dense coding feature extraction and multi-scale fusion primarily extracts features from the input remote sensing image using a Link-Net backbone network architecture through residual coding. Simultaneously, dense feature enhancement is used to directly transfer shallow features to deeper layers, reducing information loss and enhancing expressive power. A multi-scale fusion module is introduced at the network end to capture multi-level context, integrating global and local information to enhance details. The double-cross bar attention module enhances the response of the encoded features to bar features in four directions (horizontal, vertical, and double diagonal) based on multi-directional bar convolution and cross-correlation, establishing correlations between intersecting channels and optimizing feature selection accuracy. Multi-level supervised prediction based on morphological structure perception uses skip-connect fusion optimized features with deep features produced by the decoder. The prediction results are constrained by segmentation base loss, structural continuity loss, and boundary accuracy loss. Based on multi-level supervision, automatic road extraction is completed, greatly improving the accuracy and robustness of road recognition in complex scenarios. The technical architecture is as follows: Figure 2 As shown. Specifically, the method of the present invention is implemented through the following steps:

[0076] S1. Dense Encoding Feature Extraction and Multi-Scale Fusion:

[0077] S11. Acquire remote sensing images of the road area to be extracted;

[0078] S12. Dense Encoding Feature Extraction consists of residual encoding and dense feature enhancement. Residual encoding, as the "skeleton" of the module, solves the gradient vanishing problem in deep network training through skip connections, ensuring that low-level detailed features (such as road edges and small branches) can be stably propagated to high-level features, providing reliable basic features for dense feature enhancement. Dense feature enhancement, as the "enhancer" of residual encoding, focuses on feature reuse and information preservation in the encoding stage. It superimposes a dense connection mechanism on the output of the residual unit, maximizing the utilization of features at each level and further exploring the correlation between features. This achieves the dual guarantee of "residual propagation + dense reuse," solving the problem of loss of detailed information (such as boundaries and small branches) due to increased network depth in road extraction. It is particularly suitable for capturing the slender structure and complex topology of roads.

[0079] Dense feature enhancement is not independent of residual coding, but rather embedded in the concatenation of residual units, forming an alternating structure of "residual unit → dense connection → residual unit". First, the input feature F is enhanced using a single residual coding unit. in Perform a preliminary transformation and output F. res1 Then, F resi As the input to a dense convolutional unit (DCU), the input to each subsequent DCU is the concatenation of the outputs of all preceding DCUs and the outputs of preceding residual units, thus:

[0080] DCU i+1 =Concat(F re DCU i )

[0081] Among them, DCU i+1 This represents the output of the (i+1)th dense convolutional unit (DCU), ultimately yielding multi-resolution, multi-semantic hierarchical features.

[0082] The densely enhanced features are compressed in dimension by a 1×1 convolution and then input into the next residual encoding unit, and so on.

[0083] S13. The multi-scale information fusion module, located at the end of the network, is responsible for deeply integrating the multi-resolution and multi-semantic hierarchical features output from the encoding stage. This addresses the representational differences of roads at different scales (e.g., wide main roads versus narrow sidewalks, long continuous roads versus local break areas), improving the accuracy and efficiency of attribute representation. Through multi-scale feature fusion and local and global attention, this module can better capture multi-scale contextual information, enhance segmentation details, improve the clarity of segmentation results, and increase segmentation accuracy. Its structure and implementation are as follows: Figure 3 As shown.

[0084] First, the features in the encoding stage are processed using different convolutional kernels. A 3×3 convolution captures local details, preserving road boundaries and small branches; a 5×5 convolution captures the continuity features of road segments; and a 7×7 convolution captures global topological features (such as the spatial distribution of roads and buildings). The convolutional results are then summed and fused to integrate features at different scales. Simultaneously, average pooling and max pooling are used to aggregate spatial features, and local attention is obtained through a 7×7 convolution and a sigmoid activation function. Next, the spatial dimension of the feature map is compressed to 1×1, and global average pooling is performed to extract global information. An attention map is then generated using a 1×1 convolution, batch normalization, and a ReLU activation function, and its size is increased to match the dimension of the input feature map. Finally, the results of global and local attention are integrated to obtain the final output feature map.

[0085] S2. Double Crossbar Attention Optimization:

[0086] This invention constructs a double-cross strip-shaped attention mechanism. Based on multi-directional strip convolution and cross-correlation, it enhances the response to strip-shaped features in four directions: horizontal, vertical, and double diagonal. It establishes correlations between intersecting channels, highlights the consistency of road features across different channels, suppresses interference from non-road areas, and focuses on the continuous structure of the target direction. The double-cross strip-shaped attention mechanism constructs an attention system with direction selectivity and channel correlation through multi-directional feature extraction, cross-correlation modeling, and attention weight generation. The specific structure is as follows: Figure 4 As shown.

[0087] S21. Feature Extraction via Multi-directional Strip Convolution: Four types of strip convolution kernels are used: horizontal, vertical, 45°, and 135° diagonal, to capture road features in different directions. The horizontal strip convolution kernel (H) has a size of 1×k (e.g., 1×7) and slides only in the horizontal direction, responding to horizontally extending roads (e.g., east-west highways); the vertical strip convolution kernel (V) has a size of k×1 (e.g., 7×1) and responds to vertically extending roads (e.g., north-south streets); the 45° diagonal strip convolution kernel (D1) has a size of k×k, with weights concentrated in the 45° diagonal direction, responding to diagonal roads; and the 135° diagonal strip convolution kernel (D2) has a size of k×k, with weights concentrated in the 135° diagonal direction, responding to diagonally opposite roads. For the input feature map... C represents the number of channels, and H and W represent the height and width, respectively. After four types of strip convolution, feature maps in four directions are obtained:

[0088] F H =Conv H (F),F V =Conv V (F),FD1 =Conv D1 (F),F D2 =Conv D2 (F)

[0089] Among them, F H Conv represents the horizontal feature map output after horizontal strip convolution. H (F) represents a horizontal convolution operation on the feature map F, where F V Conv represents the feature map in the vertical direction output after vertical strip convolution. V (F) denotes a vertical convolution operation on the feature map F, where F D1 Conv represents the feature map output along the 45° diagonal direction after a linear convolution with a 45° diagonal line. D1 (F) indicates that a 45° diagonal convolution operation is performed on the feature map F, where F D2 Conv represents the feature map output along the 135° diagonal direction after a linear convolution along the 135° diagonal. D2 (F) indicates that a 135° diagonal convolution operation is performed on the feature map F.

[0090] S22. Cross-correlation channel association modeling: Road features exhibit directional consistency across different channels (e.g., the horizontal features of one channel are highly correlated with the horizontal features of another channel). Cross-correlation is used to quantify the strength of directional associations between channels. For each direction's feature map (e.g., F... H This process compresses the spatial dimension into a vector, resulting in the channel feature vector. (where c is the channel index), then the cross-correlation between any two channels c and c′ is:

[0091]

[0092] Among them, Corr H (c,c′) represents the cross correlation between any two channels c and c′ in the feature map in the horizontal direction; This represents the feature vector of channel c; represents the feature vector of channel c′; ‖·‖ represents the norm; · is the inner product, and the result ranges from [-1,1]. The larger the value, the higher the consistency of the features of the two channels in the horizontal direction.

[0093] Similarly, we can obtain

[0094] Vertical feature map F V Feature diagram F along the 45° diagonal direction D1 Feature diagram F along the 135° diagonal direction D2The spatial dimension is compressed into vectors, and the cross-correlation of any two channels c and c′ is calculated to obtain the cross-correlation Corr of any two channels c and c′ in the vertical feature map. V The cross-correlation between any two channels c and c′ in the feature map along the 45° diagonal direction (c, c′) Corr D1 The cross-correlation between any two channels c and c′ in the feature map along the 135° diagonal direction (c, c′) Corr D2 (c,c′), taking the average of the correlation matrices in these four directions, yields the comprehensive channel correlation matrix M(c,c′):

[0095]

[0096] The larger M(c,c′) is, the stronger the feature consistency between channels c and c′ in the four directions, and they may jointly encode road features.

[0097] S23. Combining multi-directional features and cross-channel correlations, the final attention weights are generated, consisting of two parts: channel attention and spatial attention. Based on the channel correlation matrix, the importance of each channel can be calculated, thus obtaining the channel attention.

[0098]

[0099] Where CA(c) represents the attention for channel c, M(c,c′) represents the correlation matrix between channels c and c′, σ is the Sigmoid function, and W c v is a learnable parameter c CA(c) is the channel vector of the original feature map. The higher the CA(c) value, the more significant the road features contained in the channel.

[0100] The feature maps from the four directions are weighted and fused to generate a spatial attention map:

[0101] SA(x,y)=σ(W s ·(F H (x,y)+F V (x,y)+F D1 (x,y)+F D2 (x,y)))

[0102] Where (x, y) are spatial coordinates, W s As a learnable parameter, σ outputs a weight of [0,1]. A higher value indicates a greater likelihood that the location belongs to a road.

[0103] The final attention map is achieved by multiplying channel attention and spatial attention: Attention(c,x,y)=CA(c)·SA(x,y).

[0104] By employing four types of strip convolution to accurately capture the multi-directional continuity of roads, the problem of insufficient response of traditional convolution to slender structures is solved. Cross-correlation is used to enhance the consistency of road features between channels and suppress the interference of background channels (such as the low correlation between building channels and road channels being weakened). Therefore, the double cross-strip attention model can focus more on the continuous road features in the target direction and maintain high extraction accuracy even in scenes with tree shadows and complex backgrounds.

[0105] S3. Multi-level supervised prediction based on morphological structure perception:

[0106] This invention addresses key issues in road extraction loss function design, such as structural continuity and boundary accuracy. It achieves automatic road extraction by constraining the rationality of prediction results through morphological structure perception. Furthermore, to effectively supervise the decoder's progressive generation, this invention includes intermediate supervision at each level, ensuring that each stage of the network contributes to the final segmentation and extraction results, leading to more refined and accurate output.

[0107] Morphological structure perception is derived by combining structural continuity loss, boundary accuracy loss, and base loss. Structural continuity loss quantifies the deviations between the predicted road and the real road in terms of skeleton shape, connectivity, and geometric dimensions, forcing the model to generate a continuous, unbroken road structure. Boundary accuracy loss enhances the model's attention to road edge pixels, reducing the spatial deviation between the predicted and real boundaries and ensuring a clear and distinguishable boundary between the road and the background. The base loss measures the overlap between the predicted and real roads based on the intersection-union ratio (IoU). Y represents the true label, and P represents the predicted probability map. The overall objective function is as follows:

[0108] L=λ base L base +λ struct L struct +λ bound L bound

[0109] L base L struct and L bound These represent the basic loss, structural continuity loss, and boundary accuracy loss, respectively, which are combined to form a multi-objective constraint, λ. base , λ struct and λ boudn These are the weights corresponding to the respective losses, with values ​​of 0.5, 0.2, and 0.3 in this case. This design allows the model to accurately distinguish between roads and background while maintaining structural integrity and boundary clarity, significantly improving road extraction performance in complex scenes.

[0110] (1) Structural continuity loss design

[0111] The structural continuity loss is a weighted combination of three aspects: skeleton consistency, regional connectivity, and geometric dimensions, targeting continuity constraints in different dimensions.

[0112] L struct =L sk +L cn +L sp

[0113] The road skeleton (centerline) is the core carrier of continuity. By comparing the spatial distance between the predicted skeleton and the actual skeleton, it can be ensured that the centerline does not shift or break. Therefore, distance transformation is used to perform morphological refinement on the actual label Y and the binarized predicted probability map P, respectively, to obtain the actual skeleton S. Y and predicted skeleton S P (These are all binary images, where 1 represents a skeleton pixel) After that, we have:

[0114]

[0115] Wherein d(p,S) P Let |p| be the Euclidean distance from pixel p to the nearest pixel in set S, and |S| be the total number of skeleton pixels. This is used to penalize skeleton offset (such as road center prediction deviation) and breaks (such as S). P Missing S Y (Key skeleton points in the middle).

[0116] Road breaks lead to an increase in connected components. By constraining the number and size of connected components, the "fragmented" regions in the prediction results can be reduced. Therefore, connected component analysis is performed on the binarized predicted road P, resulting in K connected domains C1, C2, ..., CK. K Its area is A1, A2, ..., A K Furthermore, the number of actual connected components K is obtained by analyzing the connected components of the real road Y. * and average area Then we have:

[0117]

[0118] Wherein, λ1 and λ2 are weights (usually λ1>λ2, prioritizing the constraint on the number of fractures), thereby penalizing excessive fractures or small fragments (with an average area much smaller than the true value).

[0119] Geometric dimensions are primarily used to constrain road width consistency and orientation smoothness (e.g., rural roads have small width fluctuations, while urban roads have regular orientations). Width consistency is obtained by calculating the deviation between the predicted width (distance perpendicular to the framework) at each point on the road and the actual width; orientation smoothness is measured by the change in the orientation angle of the framework points, penalizing sharp turns (non-intersection scenarios), thus:

[0120]

[0121] Among them, w P (i) and w Y (i) represents the predicted width and the actual road width at skeleton point i, respectively, and θ represents the width of the predicted road and the actual road width at skeleton point i, respectively. i Let be the tangent direction angle of skeleton point i.

[0122] (2) Design of Boundary Accuracy Loss

[0123] Boundary accuracy loss focuses on pixel-level errors in the road edge region, combining boundary distance metrics and edge attention mechanisms.

[0124] L bound =L hd +L ed

[0125] Boundary distance loss is applied to the ground truth label Y and the predicted probability map P using the Canny edge detection operator to obtain the ground truth boundary B. Y and prediction boundary B P And by using the average Hausdorff distance to directly quantify the degree of boundary offset, and forcing the predicted boundary to move closer to the true boundary, we have:

[0126]

[0127] Where d(x,B) is the shortest Euclidean distance from pixel p to the boundary set B.

[0128] Edge attention loss assigns higher weights to boundary pixels, enhancing the model's focus on edge regions (boundary pixels have a low proportion in the overall picture but a significant impact on visual integrity). This is achieved through the study of true boundaries (B). Y Generate a weighted graph W(x), and introduce this weight into the cross-entropy, then we have:

[0129]

[0130] Where y(x) is the true label of pixel x, and p(x) is the predicted probability, in order to improve the classification accuracy of boundary regions and avoid edge pixels being "diluted" by background pixels. This embodiment relies on the CHN6-CUG road dataset to train, validate, and evaluate the performance of the established model, exploring the real application value of joint detection and recognition technology in enhancing the comprehensive effectiveness of traffic networks. Figure 5 As shown, Figure 5(a) in the image is four high-resolution satellite remote sensing images randomly selected from the CHN6-CUG dataset; (b) is the ground truth of manually labeled roads corresponding to the four high-resolution satellite remote sensing images in (a). The remote sensing images are from Google Earth. Based on road surface coverage, the labeled roads include those with and without rail coverage. Based on geographical factors, the labeled roads include railways, highways, urban roads, and rural roads. CHN6-CUG contains 4511 labeled images of size 512×512, divided into 3608 images for model training and 903 images for testing and result evaluation. Each pixel represents an actual geographical area of ​​0.5 meters.

[0131] To improve the efficiency of image feature extraction and significantly enhance the model's generalization ability and the reliability of prediction results, various data augmentation techniques were employed to perform horizontal mirroring, vertical flipping, and 90° rotation on the images to create a road prediction dataset. By comparing the differences between the transformed images and the original images, as well as their fit with the actual detection images, the most suitable prediction results for the detection images were identified. Based on the Ubuntu 18.04 operating system environment, various neural network models were created using the PyTorch deep learning framework. To improve algorithm performance, high-performance computing tools such as CUDA and cuDNN were used to achieve efficient resource scheduling and utilization. For hardware selection, an NVIDIA GeForce GTX3080 GPU with 12GB of video memory was chosen, leveraging its excellent parallel processing capabilities to complete model training and testing. Regarding hyperparameter settings, the batch size was set to 8, and the initial learning rate was set to 0.001. During iteration, if the verification loss did not fall below the historical minimum value within four consecutive training epochs, the learning rate was adjusted to one-fifth of its original value to accelerate convergence. When the adjusted learning rate was less than 5*10... -7 The model training process will automatically end if the loss after seven consecutive iterations is still higher than the current optimal value.

[0132] The experimental quantitative evaluation method uses overall precision (OA), recall (RE), precision (PR), F1 score, and intersection-over-union ratio (IOU) as evaluation metrics. Precision describes the accuracy of the overall results, recall describes the completeness of the extracted roads, precision describes the correctness of the extracted roads, the F1 score considers both precision and recall and is a weighted harmonic average of the two, and the intersection-over-union ratio measures the degree of overlap between the predicted and actual regions. These metrics can be calculated using a confusion matrix, the definition of which is shown in Table 1.

[0133] Table 1 Definition of Confusion Matrix

[0134]

[0135] True (TP) represents the number of pixels that are detected as roads and are actually roads; True Negative (TN) represents the number of pixels that are detected as non-roads and are actually non-roads; False Positive (FP) represents the number of pixels that are detected as roads but are actually non-roads; False Negative (FN) represents the number of pixels that are detected as non-roads but are actually roads.

[0136] The evaluation indicators are obtained based on the confusion matrix as follows:

[0137] Overall accuracy: The percentage of pixels correctly classified across all categories out of the total number of pixels;

[0138]

[0139] Precision: The proportion of pixels correctly predicted as positive out of the total number of predicted positive pixels. It represents the probability that a sample predicted as positive is actually positive, and reflects the model's ability to distinguish negative samples.

[0140]

[0141] Recall: The proportion of pixels correctly predicted as positive out of the total number of actual positive pixels, that is, the probability of a positive sample being correctly detected, which reflects the model's ability to distinguish positive samples;

[0142]

[0143] F1 score: The harmonic mean of accuracy and recall, used to judge the robustness of a model;

[0144]

[0145] Intersection-Union Ratio (IU): Calculates the ratio of the intersection to the union of two sets (the prediction set and the truth set).

[0146]

[0147] (1) To better evaluate the effectiveness of the method proposed in this paper, it is compared with mainstream methods, including DANet, ABCNet, BANet, Segmenter, and UNetformer, as shown in the table. For fairness, all training parameters of the method proposed in this paper and other methods are kept consistent. As can be seen from the table below, the method proposed in this paper is significantly better than other methods.

[0148] Table 2 Comparison and analysis of accuracy evaluation of different methods

[0149] method IoU (%) OA (%) PR (%) RE(%) F1(％) DANet 88.66 99.0 85.67 88.89 86.11 ABCNet 90.72 99.26 88.54 90.31 88.72 BANet 90.84 99.08 88.68 92.45 89.40 Segmenter 94.68 99.63 93.74 95.63 94.09 UNetformer 95.04 99.63 94.03 95.96 94.66 Method of the present invention 97.38 99.79 97.01 97.84 97.29

[0150] Figure 6For the visualization comparison of the method of this invention with other methods, it can be seen from the figure that the method of this invention has better completeness in both recall and extraction of roads, especially for areas where the boundaries are not obvious. Combined with the analysis of relevant data in the figure, this invention can enhance the completeness and continuity of road selection, verifying its excellent performance in road extraction work, and maintaining a very close consistency with various measurement standards. (2) Based on the attention mechanism and combined with the self-attention mechanism, this invention proposes a double cross strip attention model to explore the long-term association characteristics of road nodes in the horizontal, vertical, and diagonal directions (from the upper left to the lower right, or from the upper right to the lower left). Through in-depth research and verification, it judges and tests whether the double cross strip attention mechanism is suitable for use in road extraction tasks, and the specific performance of this attention. To this end, based on the proposed architecture, the strip attention module is compared with other attention modules through a maximum of 100 iterations, including SE channel attention, CBAM channel spatial attention, DA dual attention, CCA channel cross attention and the improved CCA module.

[0151] As can be seen from the above, when adding other attention modules besides SE and DA, the training loss of the network continues to decrease, and the learning curve becomes smoother. Figure 7 Analysis reveals that the SE attention mechanism significantly improves the accuracy of road recognition. However, using it alone is difficult to effectively extract road features and may even affect the model's generalization ability. Cognitive processing methods using non-local self-attention mechanisms encounter numerous challenges in road extraction due to their computational complexity. CBAM, CCA, and dCCA all effectively improve the performance of road surface information extraction models. CCA stands out, with a 1.02% increase in Intersection over Union (IoU) compared to the traditional baseline model without attention mechanisms. This demonstrates that local attention strategies are better than global attention strategies, significantly improving the accuracy and speed of road surface feature recognition. Compared to existing attention designs, the double-cross strip attention mechanism used in this invention performs even better, improving the F1 score and IoU by 1.62% and 1.41%, respectively, clearly demonstrating its superiority in improving road surface information processing. This model can comprehensively cover long-distance pixel-related features in the horizontal, vertical, and two diagonal directions, exhibiting significant technical advantages in the road extraction stage.

[0152] Table 3 Comparison of training results with different attention modules inserted.

[0153] Attention module Maximum number of iterations Actual number of iterations Loss upon termination none 100 80 0.0562 SE channel attention 100 82 0.1051 DA dual attention 100 99 0.1823 CBAM Spatial Channel Attention 100 100 0.0607 CCA channel cross attention 100 76 0.0562 Improved CCA 100 99 0.0391 Double cross bar attention 100 89 0.0532

[0154] Table 4 Comparison of road extraction performance with different attention modules

[0155] Attention module ACC PR RE F1 IoU none 0.9509 0.7903 0.6994 0.7249 0.5684 SE channel attention 0.9707 0.7849 0.6329 0.7121 0.5529 DA dual attention 0.9591 0.8359 0.3566 0.4999 0.3333 CBAM Spatial Channel Attention 0.9709 0.7907 0.6702 0.7255 0.5692 CCA channel cross attention 0.9704 0.7589 0.7089 0.7331 0.5786 Improved CCA 0.9707 0.7888 0.6736 0.7267 0.5707 Double cross bar attention 0.9753 0.7619 0.7121 0.7361 0.5825

[0156] (3) The table below shows the quantitative evaluation results of different modules. As can be seen from the table, the double cross bar attention fusion baseline proposed in this invention performs well in all indicators on the dataset. Subsequently, multiple modules proposed in this invention were verified. The double cross bar attention module improved the IoU on the dataset by 0.63%. After adding the multi-scale fusion module, the IoU on the dataset increased significantly. The main reason may be that the types of roads in the dataset are more complex and multi-scale information needs to be paid attention to. After introducing multi-level supervision, the IoU of the dataset also improved by 0.86%.

[0157] Table 5 Experimental results for different modules

[0158] Network module IoU (%) Overall accuracy (%) Recall rate (%) Accuracy (%) F1(％) Baseline 96.52 99.75 96.24 96.89 96.25 Multi-scale fusion 97.11 99.77 96.62 97.79 97.09 +Double cross-shaped attention 97.15 99.78 96.91 97.81 97.12 Multi-level supervision 97.38 99.79 97.01 97.84 97.29

Claims

1. A road intelligent extraction method based on occlusion perception, characterized in that, The method includes the following steps: S1. Dense Coding Feature Extraction and Multi-Scale Fusion: Acquire remote sensing images of the road area to be extracted, and combine residual coding and dense convolutional units based on the Link-Net backbone network architecture. This approach directly transmits shallow features to deep layers, forces feature reuse, reduces information loss, and enhances expressive power. Furthermore, a multi-scale fusion module is introduced at the end of the network to obtain a feature map that integrates global and local information through multi-level deep analysis. S2. Double Cross-Strip Attention Optimization: The feature map obtained in step S1 is processed using horizontal, vertical, and double diagonal strip convolution operations to establish the correlation between intersecting channels. Combining multi-directional features and intersecting channel correlations, attention weights are generated to obtain optimized features. The double cross-strip attention optimization specifically includes the following sub-steps: S21. Feature extraction using multi-directional strip convolution: The feature map output from step S1 is processed using four types of strip convolution kernels: horizontal, vertical, 45° diagonal, and 135° diagonal, to capture road features in different directions. The size of the horizontal strip convolution kernel H is 1×k, where k represents the receptive field size and ranges from [3,15]. It slides only in the horizontal direction to respond to horizontally extending roads. The vertical strip convolution kernel V has a size of k×1, responding to vertically extending roads; the 45° diagonal strip convolution kernel D1 has a size of k×k, with weights concentrated in the 45° diagonal direction, responding to diagonal roads; the 135° diagonal strip convolution kernel D2 has a size of k×k, with weights concentrated in the 135° diagonal direction, responding to reverse diagonal roads. For the input feature map C represents the number of channels, and H and W represent the height and width, respectively. After four types of strip convolution, feature maps in four directions are obtained: ， in, This represents the horizontal feature map output after horizontal strip convolution. This indicates that a horizontal convolution operation is performed on the feature map F. This represents the feature map in the vertical direction output after vertical strip convolution. This indicates that a vertical convolution operation is performed on the feature map F. This represents the feature map output along the 45° diagonal direction after a 45° diagonal linear convolution. This indicates that a 45° diagonal convolution operation is performed on the feature map F. This represents the feature map output along the 135° diagonal direction after a linear convolution with a 135° diagonal line. This indicates that a 135° diagonal convolution operation is performed on the feature map F; S22. Channel correlation modeling of cross-correlation: The horizontal feature map output after performing horizontal strip convolution on step S21. By compressing the spatial dimension into a vector, we obtain the channel feature vector. ,in If it is a channel index, then any two channels , The cross-correlation is: ， in, Representing any two channels in the feature map in the horizontal direction , Cross-correlation; Indicates channel eigenvectors; Indicates channel eigenvectors; Represents the norm; The inner product has a result range of [-1, 1]. A larger value indicates a higher consistency of features between the two channels in the horizontal direction. Similarly, the feature map in the vertical direction obtained in step S21... Feature map along the 45° diagonal direction Feature map along the 135° diagonal direction Compress the spatial dimension into a vector and perform arbitrary two-channel processing. , Cross-correlation calculation yields any two channels in the vertical feature map. , Cross-correlation Any two channels in the feature map along the 45° diagonal direction , Cross-correlation Any two channels in the feature map along the 135° diagonal direction , Cross-correlation The average of the correlation matrices in these four directions yields the comprehensive channel correlation matrix. : ， in, The larger the channel, the better. and The stronger the feature consistency in the four directions, the more likely they are to jointly encode road features; S23. Based on the integrated channel correlation matrix obtained in step S22 Calculate the importance of each channel and obtain channel attention: ， in, Indicates channel attention, Indicates channel and The correlation matrix between them For the Sigmoid function, For learnable parameters, This represents the channel vectors of the original feature map. The higher the value, the more significant the road features contained in the passage; The feature maps from the four directions are weighted and fused to generate a spatial attention map: ， in For spatial coordinates, For learnable parameters, Output a weight of [0,1], where a higher value indicates a greater likelihood that the location belongs to a road; The final attention map is achieved by multiplying channel attention and spatial attention: ； S3. Multi-level supervised prediction for morphological structure perception: Deep features are generated by the decoder, and the optimized features are fused with the deep features by using skip connections. The rationality of the morphological structure perception constraint prediction results obtained by combining the segmentation base loss, structural continuity loss and boundary accuracy loss is evaluated, and automatic road extraction is completed based on multi-level supervision.

2. The road intelligent extraction method based on occlusion perception according to claim 1, characterized in that, Step S1, which involves dense coding feature extraction and multi-scale fusion, specifically includes the following sub-steps: S11. Acquire remote sensing images of the road area to be extracted; S12. Based on the Link-Net backbone network architecture, residual coding and dense feature enhancement are combined to directly pass shallow features to deep layers, forcing feature reuse. Specifically, dense convolutional units... In the process of embedding residual coding units in a concatenation, a "residual coding unit → dense convolutional unit" is formed. → The alternating structure of "residual coding units" first processes the input features through one residual coding unit. Perform preliminary transformation and output. Then, As a dense convolutional unit The input, and each subsequent one The input is the previous one. The output is concatenated with the output of the previous residual unit, resulting in: ， in, Indicates the first Sub-dense convolutional units The output of the function ultimately yields multi-resolution, multi-semantic hierarchical features. The densely enhanced features are compressed in dimension by a 1×1 convolution and then input into the next residual encoding unit, and so on. S13. The multi-scale information fusion module is located at the end of the network and is responsible for the deep integration of multi-resolution and multi-semantic level features output from the encoding stage. Specifically: S131. The multi-resolution and multi-semantic hierarchical features output during the encoding stage are processed using different convolution kernels. Local details are captured by 3×3 convolution, continuous features of road segments are captured by 5×5 convolution, and global topological features are captured by 7×7 convolution. The convolution results are then summed and fused to integrate features at different scales. S132. Aggregate the fused features of different scales obtained in S131 using average pooling and max pooling, and obtain local attention through 7×7 convolution and sigmoid activation function; S133. The spatial dimension of the multi-resolution, multi-semantic hierarchical features output in the encoding stage is compressed to 1×1, global average pooling is performed to extract global information; and attention maps are generated using 1×1 convolution, batch normalization and ReLU activation function, and then their size is expanded to match the dimension of the input feature map to obtain global attention. S134. Integrate the results of global attention and local attention to obtain the final output feature map.

3. The road intelligent extraction method based on occlusion perception according to claim 1, characterized in that, Step S3 describes a multi-level supervised prediction based on morphological structure perception. Morphological structure perception is obtained by combining structural continuity loss, boundary accuracy loss, and basic loss. Structural continuity loss quantifies the deviations between the predicted road and the real road in terms of skeleton shape, connectivity, and geometric dimensions, forcing the model to generate a continuous, unbroken road structure. Boundary accuracy loss enhances the model's focus on road edge pixels, reducing the spatial deviation between the predicted boundary and the real boundary, ensuring a clear and distinguishable boundary between the road and the background. The basic loss... The intersection-overlap ratio (IoU) measures the degree of overlap between the predicted and actual roads. , For real labels, To predict the probability map, the overall objective function is... as follows: ， , and These represent the basic loss, structural continuity loss, and boundary accuracy loss, respectively, which are combined to form multi-objective constraints. , and These are the weights corresponding to the basic loss, structural continuity loss, and boundary accuracy loss, respectively; the specific design methods for the basic loss, structural continuity loss, and boundary accuracy loss are as follows: (1) Structural continuity loss design: The structural continuity loss is a weighted combination of three aspects: skeleton consistency, regional connectivity, and geometric dimensions, addressing continuity constraints in different dimensions: ， in, This indicates a loss of skeleton consistency. Indicates the loss of regional connectivity. Indicates geometric loss; The road framework, i.e., the central axis of the road, is the core carrier of continuity. Distance transformation is used to apply it to the real labels. And the binarized predicted road probability map Morphological refinement was performed to obtain the true skeleton. and predict skeleton Then, there is: ， in, For pixels to set The Euclidean distance to the nearest pixel in the middle. The total number of skeleton pixels is used to penalize skeleton offset and breakage; Binarized prediction road probability map Perform connected component analysis to obtain Connected components Its area is ; and the real road The connected component analysis yields the actual number of connected components. and average area Then we have: ， in, , As weight and Prioritize limiting the number of fractures to punish excessive fractures or small fragments. Geometric dimensions are used to constrain road width consistency and orientation smoothness. Width consistency is obtained by calculating the deviation between the predicted width at each point on the road and the actual width, i.e., the distance perpendicular to the skeleton. Orientation smoothness is measured by the change in the orientation angle of the skeleton points, penalizing sharp turns in non-intersection scenarios. Therefore: ， in, and The predicted results and the actual road at the skeleton points are respectively. Width at that point skeleton points The tangent direction angle; (2) Design of boundary accuracy loss Boundary accuracy loss focuses on pixel-level errors in the road edge region, combining boundary distance metrics and edge attention mechanisms: ， in, Indicates the boundary distance loss. This indicates peripheral attention loss; Boundary distance loss is applied to the ground truth labels using the Canny edge detection operator. and prediction probability map Processing is performed to obtain the true boundaries. and prediction boundary And by using the average Hausdorff distance to directly quantify the degree of boundary offset, and forcing the predicted boundary to move closer to the true boundary, we have: ， in, For pixels To the boundary set The shortest Euclidean distance; Through real boundaries Generate weight graph Introducing this weight into the cross-entropy framework, we have: ， in, For pixels The true label, To predict probabilities.

Images