A method and system for semantic segmentation of highway images

Abstract

The application provides a highway image semantic segmentation method and system, relates to the technical field of image processing, and comprises the following steps: performing size normalization and pixel normalization on a highway image to obtain an input image; inputting the input image into a lightweight feature extraction network to extract multi-scale features, including shallow features, middle features, high features and deep features; inputting the deep features into a road strip context enhancement network to enhance the long strip continuous structure of the highway by using horizontal strip context, vertical strip context, local convolution features and global semantic features, and obtain road enhancement features; and performing multi-scale fusion decoding on the road enhancement features, the high features, the middle features and boundary enhancement features to recover the spatial resolution step by step and generate a final semantic segmentation result. The application can improve the segmentation accuracy and boundary integrity of the road area, emergency lane, lane line, vehicle, guardrail and shoulder in the highway image of the unmanned aerial vehicle.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, specifically relating to a semantic segmentation method and system for highway images. Background Technology

[0002] With the development of drone inspection technology and intelligent transportation systems, using drones for highway inspection, monitoring, and scene understanding has become an important technological direction in traffic management. Drones can quickly acquire large-scale highway images from an aerial perspective, providing data support for road condition analysis, anomaly detection, traffic violation assistance, and infrastructure inspection.

[0003] Semantic segmentation technology can classify each pixel in an image, providing more refined region-level information compared to object detection methods. Therefore, in highway scenarios, semantic segmentation can be used to identify traffic elements such as road areas, emergency lanes, lane lines, vehicles, guardrails, median strips, shoulders, and obstacles, providing a foundation for subsequent traffic incident analysis and decision support.

[0004] However, semantic segmentation of highway images by drones still faces the following problems:

[0005] First, in highway scenarios, objects such as roads, emergency lanes, guardrails, central dividers, and shoulders are usually long, continuous, and directional, and the local receptive field of ordinary convolutional networks is difficult to fully model such long-distance continuous structures.

[0006] Second, from the perspective of drone aerial photography, fine-grained structures such as lane lines, emergency lane boundaries, guardrail edges and vehicle outlines are relatively narrow, and the boundaries are prone to blurring or breaking during feature downsampling and decoding.

[0007] Third, the vehicle scale varies greatly in highway images, and distant small targets and road structures are easily affected by changes in lighting, shadows, occlusion, and shooting angle, leading to unstable segmentation results.

[0008] Fourth, some high-precision semantic segmentation models have a large computational load, which is not conducive to rapid deployment in UAV inspections or edge computing devices.

[0009] Therefore, there is a need to provide a semantic segmentation method for UAV highway images that can enhance the ability to express the long road structure and recover fine-grained boundaries of highways while maintaining high computational efficiency. Summary of the Invention

[0010] The purpose of this invention is to provide a semantic segmentation method and system for highway images to improve the aforementioned problems. To achieve this objective, the technical solution adopted by this invention is as follows:

[0011] In a first aspect, this application provides a semantic segmentation method for highway images, comprising:

[0012] Aerial images of highways are acquired, and the images are normalized in size and pixels to obtain the input image.

[0013] The input image is fed into a lightweight feature extraction network to extract multi-scale features, which include shallow features, mid-level features, high-level features, and deep features.

[0014] The deep features are input into the road strip context enhancement network, and the long strip continuous structure of the highway is enhanced by using horizontal strip context, vertical strip context, local convolutional features and global semantic features to obtain road enhancement features;

[0015] The shallow and middle layer features are input into the boundary refinement network to generate a boundary attention map. The initial boundary features are then weighted and augmented based on the boundary attention map to obtain the enhanced boundary features.

[0016] The road enhancement features, high-level features, mid-level features, and boundary enhancement features are input into a multi-scale fusion decoder to obtain the decoded features;

[0017] The decoded features are subjected to classification convolution and upsampling processing to obtain the semantic segmentation results of the highway image.

[0018] Secondly, this application also provides a semantic segmentation system for highway images, comprising:

[0019] The first module is used to acquire aerial images of highways and perform size normalization and pixel normalization processing on the aerial images of highways to obtain an input image.

[0020] The second module is used to input the input image into a lightweight feature extraction network to extract multi-scale features, wherein the multi-scale feature map includes shallow features, mid-level features, high-level features and deep features;

[0021] The third module is used to input the deep features into the road strip context enhancement network, and to enhance the long strip continuous structure of the highway by using horizontal strip context, vertical strip context, local convolutional features and global semantic features to obtain road enhancement features;

[0022] The fourth module is used to input the shallow features and the middle features into the boundary refinement network to generate a boundary attention map, and to perform weighted and residual enhancement on the initial boundary features based on the boundary attention map to obtain the boundary enhanced features;

[0023] The fifth module is used to input the road enhancement features, high-level features, mid-level features, and boundary enhancement features into the multi-scale fusion decoder to obtain the decoded features;

[0024] The sixth module is used to perform classification convolution and upsampling processing on the decoded features to obtain the semantic segmentation results of the highway image.

[0025] The beneficial effects of this invention are as follows:

[0026] This invention improves the segmentation capability of continuous areas such as roads, emergency lanes, guardrails, medians, and shoulders by jointly modeling long, continuous structures in highway scenes through horizontal average pooling, vertical average pooling, local convolution, and global average pooling branches. It generates boundary attention maps using spatial details from shallow features and local semantic information from mid-level features, and obtains boundary enhancement features through attention weighting and residual enhancement, improving the segmentation integrity of lane lines, emergency lane boundaries, vehicle outlines, and guardrail edges. Multi-scale fusion progressively integrates road enhancement features, high-level semantic features, mid-level structural features, and boundary enhancement features, enabling the model to possess both global road structure understanding and local boundary recovery capabilities.

[0027] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing embodiments of the invention. Attached Figure Description

[0028] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0029] Figure 1 This is a flowchart of a semantic segmentation method for highway images according to an embodiment of this application;

[0030] Figure 2 This is a schematic diagram of the road strip context enhancement process according to an embodiment of this application;

[0031] Figure 3 This is a schematic diagram illustrating the boundary refinement process of an embodiment of this application;

[0032] Figure 4 This is a schematic diagram of the multi-scale fusion decoding process in an embodiment of this application;

[0033] Figure 5 This is a structural diagram of a semantic segmentation device for highway images according to an embodiment of this application.

[0034] Symbol descriptions: 800 - Semantic segmentation device for highway images; 801 - Processor; 802 - Memory; 803 - Multimedia component; 804 - I / O interface; 805 - Communication component. Detailed Implementation

[0035] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0036] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0037] Example 1:

[0038] See Figure 1 This embodiment provides a semantic segmentation method for highway images, including steps S100, S200, S300, S400, and S500.

[0039] S100. Acquire aerial images of the highway and perform size normalization and pixel normalization processing on the aerial images of the highway to obtain the input image.

[0040] The input image is an aerial photograph of a highway taken by a drone, denoted as:

[0041]

[0042] in, Indicates the input image. Indicates the image height. 3 indicates the image width, and 3 indicates the RGB three channels.

[0043] S200. Input the input image into a lightweight feature extraction network to extract multi-scale features, including shallow features, mid-level features, high-level features and deep features;

[0044] The input image is processed by a lightweight feature extraction network to obtain feature maps at four scales. In one optional embodiment, the lightweight feature extraction backbone includes an initial convolutional layer and four feature extraction stages.

[0045] First, input an image. through Initial features are obtained by processing an initial convolutional layer with a stride of 2 and 32 output channels.

[0046] Then, the initial features go through four feature extraction stages in sequence.

[0047] The first feature extraction stage uses depthwise separable convolution for downsampling and feature extraction, outputting shallow features:

[0048] The second feature extraction stage outputs mid-level features:

[0049] The third feature extraction stage outputs high-level features:

[0050] The fourth feature extraction stage outputs deep features:

[0051] The depthwise separable convolution includes depthwise convolution and pointwise convolution. Depthwise convolution is used to extract spatial features, while pointwise convolution is used to fuse channel information. Each convolutional unit is followed by a batch normalization layer and a ReLU activation function to improve the network's training stability and non-linear expressive power.

[0052] Shallow features Primarily preserves spatial texture and boundary details in the image; mid-level features It simultaneously contains local structural and semantic information; high-level features Representing semantic information of categories such as roads, vehicles, guardrails, and medians; deep features It represents global context and overall road layout information.

[0053] S300. Input the deep features into the road strip context enhancement network, and enhance the long strip-shaped continuous structure of the highway using horizontal strip context, vertical strip context, local convolutional features, and global semantic features to obtain road enhancement features. See [link to relevant documentation]. Figure 2 ;

[0054] S310, Compress the number of channels in deep features using 1×1 convolution;

[0055] Input deep features enter Convolution performs channel compression to obtain compressed features. :

[0056]

[0057] in:

[0058]

[0059] S320. The compressed deep features are processed through four parallel branches: horizontal average pooling, vertical average pooling, local convolution, and global average pooling, to extract horizontal strip context, vertical strip context, local convolution features, and global semantic features.

[0060] The horizontal average pooling branch is used to extract horizontal stripe context information. This branch first... conduct Average pooling is used to obtain horizontal stripe features; then... The convolution adjusts the number of channels to 128 and performs BN+ReLU processing; finally, the features are restored to their original size through upsampling, broadcast expansion, or size restoration operations. Spatial dimensions yield lateral contextual features. This branch can enhance the representation of lateral road boundaries, pavement areas, and local road structures in highway images.

[0061] The vertical average pooling branch is used to extract vertical strip context information. This branch first... conduct Average pooling is used to obtain vertical strip features; then... The convolution adjusts the number of channels to 128 and performs BN+ReLU processing; finally, it is restored to its original size using a size restoration operation. Spatial dimensions yield vertical contextual features. This branch can enhance the continuous structural expression of highways along the road's extension direction, and is especially suitable for long, narrow areas such as main road surfaces, emergency lanes, guardrails, and central medians.

[0062] Local convolution branches use Convolution, BN, and ReLU pairs Perform local feature extraction and output local convolutional features. This branch is used to preserve local details such as vehicle edges, lane line morphology, road surface texture, and small obstacles, preventing local features from being weakened during strip pooling.

[0063] Global semantic branch first Perform global average pooling to obtain Global semantic features; then through The convolution adjusts the number of channels to 128 and performs BN+ReLU processing; finally, it is restored to its original size using a size restoration operation. Spatial dimensions yield global context features. This branch is used to provide overall layout priors for highway scenarios, helping the model distinguish between road areas and non-road areas.

[0064] S330. The outputs of the four branches are concatenated along the channel dimension and then fused by a 1×1 convolution to obtain the fused features;

[0065]

[0066] in, For the splicing result, For lateral context features, For vertical context features, Local convolution features For global context features; then through Convolutional fusion yields fused features. .

[0067] S340. The fused features are added to the input deep features through residual connections to obtain road enhancement features;

[0068]

[0069] in, Enhance road features; Features of fusion; For deep features; Spatial scale and number of channels Maintaining consistency, that is: ;

[0070] S400. Input the shallow and middle layer features into the boundary refinement network to generate a boundary attention map. Then, based on the boundary attention map, perform weighted and residual enhancement on the initial boundary features to obtain the enhanced boundary features. See [link to relevant documentation]. Figure 3 ;

[0071] S410. Upsample the mid-layer features to the same spatial scale as the shallow-layer features, and perform channel alignment using a 1×1 convolution.

[0072] Mid-layer features Perform a 2x upsampling to change its spatial scale from Restore to Subsequently, the number of channels was adjusted to 64 using 1×1 convolution, and then processed by Batch Normalization (BN) and ReLU (Revised Linear Unit) to obtain the mid-layer boundary features:

[0073] S420. Then, the middle layer features are concatenated with the shallow layer features processed by 3×3 convolution, and then extracted by 3×3 convolution to obtain the initial boundary features.

[0074] shallow features Inputting 3×3 convolution, BN, and ReLU, we obtain shallow boundary features: ;

[0075] The shallow boundary features and the middle boundary features are concatenated: ; Features after splicing;

[0076] The concatenated features are processed by 3×3 convolution, BN, and ReLU to obtain the initial boundary features:

[0077] S430. Perform 1×1 convolution and Sigmoid activation on the initial boundary features to generate a boundary attention map;

[0078] Initial boundary features Input a 1×1 convolution, compress the number of channels to 1, and generate a boundary attention map using a sigmoid activation function:

[0079]

[0080] in:

[0081] S440. Finally, the initial boundary features are weighted element-wise using the boundary attention map, and the boundary enhancement features are obtained using residual enhancement:

[0082]

[0083] in, This indicates element-wise multiplication. Represents the boundary attention map. Initial boundary features; boundary enhancement features Used to enhance fine-grained structures such as lane lines, emergency lane boundaries, guardrail edges, vehicle outlines, and road boundaries.

[0084] S500: Input the road enhancement features, high-level features, mid-level features, and boundary enhancement features into the multi-scale fusion decoder to obtain the decoded features. (See below) Figure 4 ;

[0085] S510. After upsampling the road enhancement features, fuse them with the high-level features to obtain the first-level decoding features;

[0086] First, enhance the road features Perform a 2x upsampling to change its spatial scale from Restore to Subsequently through The convolution adjusts the number of channels to 256 and then processes them using BN+ReLU.

[0087] Processed road enhancement features and high-level features Perform channel splicing, then through Convolution, BN, and ReLU processing are used to obtain the first-level decoded features. :

[0088]

[0089] S520. Then, the first-level decoding features are upsampled and fused with the middle-layer features to obtain the second-level decoding features.

[0090] First-level decoding features Perform a 2x upsampling to change its spatial scale from Restore to Subsequently through The convolution adjusts the number of channels to 128 and then processes them using BN+ReLU.

[0091] Processed features and mid-level features Perform channel splicing, then through Convolution, BN, and ReLU processing are used to obtain the second-level decoded features. :

[0092]

[0093] S530. The second-level decoding features are fused with the boundary enhancement features to obtain the third-level decoding features;

[0094] Second-level decoding features Perform a 2x upsampling to change its spatial scale from Restore to Subsequently, it was combined with boundary enhancement features. Channel splicing is performed, and the process is carried out using boundary-guided fusion convolution.

[0095] The boundary-guided fusion convolution employs Convolution, Batch Normalization (BN), and ReLU are used to fuse semantic decoding features with boundary enhancement features to obtain third-level decoding features. :

[0096]

[0097] S600. Perform classification convolution and upsampling processing on the decoded features to obtain the semantic segmentation result of the highway image;

[0098] S610, Perform the third-level decoding features Classification convolution yields category probability feature maps. , Indicates the number of semantic categories.

[0099] S620. Upsample the category probability feature map to restore it to the input image size;

[0100] S630. Based on the upsampled category probability feature map, select the category number with the highest probability for each pixel as its final semantic category.

[0101] The category of each pixel can be determined in the following way:

[0102]

[0103] in, Represents pixels Predicted semantic category; This represents the total number of semantic categories that the pixel may correspond to.

[0104] S640. Obtain the semantic segmentation result based on the final semantic category of all pixels. ;

[0105] In one alternative implementation, the semantic segmentation results include, but are not limited to, the following categories: background, main road surface, emergency lane, lane lines, vehicles, guardrails, central divider, shoulder, obstacles, or debris.

[0106] The above categories can be added, deleted, or merged according to specific highway inspection tasks without affecting the core technical solution of this invention.

[0107] In one alternative implementation, a joint loss function is used when training the network:

[0108]

[0109] in, For the total loss, Main segmentation loss, For the loss of border surveillance, To assist in monitoring losses, and These are the weighting coefficients. Details are as follows:

[0110] Based on the semantic segmentation results and the ground truth semantic labels (the actual semantic meaning corresponding to each pixel), calculate the cross-entropy loss and Dice loss to obtain the main segmentation loss. ;

[0111] Based on the real semantic label map in the training set (each pixel value represents a semantic category), boundary information is extracted through image processing methods (edge ​​extraction, erosion dilation difference, or contour extraction) to generate corresponding boundary supervision labels;

[0112] The boundary attention map and the boundary supervision label are compared pixel by pixel, and the difference between them is calculated using the binary cross-entropy loss function to obtain the boundary supervision loss. ;

[0113] The auxiliary supervision loss can be set on the intermediate decoded features to improve the stability of network training and the effect of gradient propagation.

[0114] The total network loss is calculated based on the main segmentation loss, boundary supervision loss, and auxiliary supervision loss. Each network is then trained based on the total network loss to obtain the semantic segmentation model.

[0115] Example 2:

[0116] This embodiment provides a semantic segmentation system for highway images, including: The first module is used to acquire aerial images of highways and perform size normalization and pixel normalization processing on the aerial images of highways to obtain an input image. The second module is used to input the input image into a lightweight feature extraction network to extract multi-scale features, including shallow features, mid-level features, high-level features and deep features; The third module is used to input the deep features into the road strip context enhancement network, and to enhance the long strip continuous structure of the highway by using horizontal strip context, vertical strip context, local convolutional features and global semantic features to obtain road enhancement features; The fourth module is used to input the shallow features and the middle features into the boundary refinement network to generate a boundary attention map, and to perform weighted and residual enhancement on the initial boundary features based on the boundary attention map to obtain the boundary enhanced features; The fifth module is used to input the road enhancement features, high-level features, mid-level features, and boundary enhancement features into the multi-scale fusion decoder to obtain the decoded features; The sixth module is used to perform classification convolution and upsampling processing on the decoded features to obtain the semantic segmentation results of the highway image.

[0117] As an optional implementation, the third module includes: The first unit is used to compress the number of channels in deep features through 1×1 convolution; The second unit is used to extract horizontal strip context, vertical strip context, local convolution features and global semantic features from the compressed deep features through four parallel branches: horizontal average pooling, vertical average pooling, local convolution and global average pooling. The third unit is used to concatenate the outputs of the four branches along the channel dimension and fuse them through a 1×1 convolution to obtain the fused features; The fourth unit is used to add the fused features to the input deep features through residual connections to obtain road enhancement features.

[0118] As an optional implementation, the fourth module includes: The fifth unit is used to upsample the mid-level features to the same spatial scale as the shallow-level features and perform channel alignment through 1×1 convolution; The sixth unit is used to concatenate the middle-layer features with the shallow-layer features processed by 3×3 convolution, and then extract the initial boundary features through 3×3 convolution. The seventh unit is used to perform 1×1 convolution and Sigmoid activation on the initial boundary features to generate a boundary attention map.

[0119] As an optional implementation, the fifth module includes: The eighth unit is used to upsample the road enhancement features and fuse them with high-level features to obtain the first-level decoded features; The ninth unit is used to further upsample the first-level decoding features and fuse them with the mid-level features to obtain the second-level decoding features; The tenth unit is used to fuse the second-level decoding features with the boundary enhancement features to obtain the third-level decoding features.

[0120] Example 3: Corresponding to the above method embodiments, this embodiment also provides a semantic segmentation device for highway images. The semantic segmentation device for highway images described below can be referred to in correspondence with the semantic segmentation method for highway images described above.

[0121] Figure 5 This is a block diagram illustrating a semantic segmentation device 800 for highway images according to an exemplary embodiment. Figure 5As shown, the semantic segmentation device 800 for highway images includes a processor 801 and a memory 802. The semantic segmentation device 800 may also include one or more of a multimedia component 803, an input / output (I / O) interface 804, and a communication component 805. The processor 801 controls the overall operation of the semantic segmentation device 800 to complete all or part of the steps in the aforementioned semantic segmentation method for highway images. The memory 802 stores various types of data to support the operation of the semantic segmentation device 800. This data may include, for example, commands for any application or method operating on the semantic segmentation device 800, and application-related data such as contact data, sent and received messages, images, audio, video, etc. The memory 802 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk.

[0122] Multimedia component 803 may include a screen and an audio component. The screen may be, for example, a touchscreen, and the audio component is used to output and / or input audio signals. For example, the audio component may include a microphone for receiving external audio signals.

[0123] The received audio signal can be further stored in memory 802 or transmitted via communication component 805. The audio component also includes at least one speaker for outputting audio signals. I / O interface 804 provides an interface between processor 801 and other interface modules, such as a keyboard, mouse, buttons, etc. These buttons can be virtual or physical. Communication component 805 is used for wired or wireless communication between the semantic segmentation device 800 of the highway image and other devices. Wireless communication, such as Wi-Fi, Bluetooth, Near Field Communication (NFC), 2G, 3G, or 4G, or a combination thereof, is used; therefore, the corresponding communication component 805 may include a Wi-Fi module, a Bluetooth module, or an NFC module.

[0124] Example 4: Corresponding to the above embodiment of the semantic segmentation method for highway images, this embodiment also provides a readable storage medium. The readable storage medium described below can be referred to in conjunction with the semantic segmentation method for highway images described above.

[0125] A readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the semantic segmentation method embodiment for highway images described above.

[0126] Specifically, the readable storage medium can be a USB flash drive, external hard drive, read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk, or any other readable storage medium capable of storing program code.

[0127] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0128] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A semantic segmentation method for highway images, characterized in that, include: Aerial images of highways are acquired, and the images are normalized in size and pixels to obtain the input image. The input image is fed into a lightweight feature extraction network to extract multi-scale features, which include shallow features, mid-level features, high-level features, and deep features. The deep features are input into the road strip context enhancement network, and the long strip continuous structure of the highway is enhanced by using horizontal strip context, vertical strip context, local convolutional features and global semantic features to obtain road enhancement features; The shallow and middle layer features are input into the boundary refinement network to generate a boundary attention map. The initial boundary features are then weighted and augmented based on the boundary attention map to obtain the enhanced boundary features. The road enhancement features, high-level features, mid-level features, and boundary enhancement features are input into a multi-scale fusion decoder to obtain the decoded features; The decoded features are subjected to classification convolution and upsampling processing to obtain the semantic segmentation results of the highway image.

2. The semantic segmentation method for highway images according to claim 1, characterized in that, The deep features are input into the road strip context enhancement network, which utilizes horizontal strip context, vertical strip context, local convolutional features, and global semantic features to enhance the long, continuous structure of the highway, including: The number of channels in deep features is compressed using 1×1 convolution; The compressed deep features are processed through four parallel branches: horizontal average pooling, vertical average pooling, local convolution, and global average pooling, to extract horizontal strip context, vertical strip context, local convolution features, and global semantic features. The outputs of the four branches are concatenated along the channel dimension and then fused using a 1×1 convolution to obtain the fused features; The fused features are added to the input deep features through residual connections to obtain road enhancement features.

3. The semantic segmentation method for highway images according to claim 1, characterized in that, The shallow and mid-level features are input into a boundary refinement network to generate a boundary attention map, including: The mid-level features are upsampled to the same spatial scale as the shallow-level features, and channel alignment is performed using 1×1 convolution. Then, the mid-layer features are concatenated with the shallow-layer features processed by 3×3 convolution, and then extracted by 3×3 convolution to obtain the initial boundary features; The initial boundary features are subjected to 1×1 convolution and Sigmoid activation to generate a boundary attention map.

4. The semantic segmentation method for highway images according to claim 1, characterized in that, include: The road enhancement features, high-level features, mid-level features, and boundary enhancement features are input into a multi-scale fusion decoder to obtain decoded features, including: The road enhancement features are upsampled and then fused with high-level features to obtain the first-level decoding features; The first-level decoding features are then upsampled and fused with the mid-level features to obtain the second-level decoding features; The second-level decoding features are fused with the boundary enhancement features to obtain the third-level decoding features.

5. The semantic segmentation method for highway images according to claim 4, characterized in that, The decoded features are subjected to classification convolution and upsampling processing to obtain the semantic segmentation results of the highway image, including: Perform a 1×1 classification convolution on the third-level decoded features to obtain the category probability feature map; Upsample the category probability feature map to restore it to the input image size; Based on the upsampled category probability feature map, the category number with the highest probability is selected for each pixel as its final semantic category; The semantic segmentation result is obtained based on the final semantic category of all pixels.

6. The semantic segmentation method for highway images according to claim 1, characterized in that, The method further includes: Based on the semantic segmentation results and the true semantic labels, calculate the cross-entropy loss and Dice loss to obtain the main segmentation loss; Based on the real semantic label map in the training set, boundary information is extracted through image processing to generate corresponding boundary supervision labels; The boundary attention map and the boundary supervision label are compared pixel by pixel, and the difference between the two is calculated using the binary cross-entropy loss function to obtain the boundary supervision loss. The total network loss is calculated based on the main segmentation loss and the boundary supervision loss. Each network is then trained based on the total network loss to obtain the semantic segmentation model.

7. A semantic segmentation system for highway images, characterized in that, include: The first module is used to acquire aerial images of highways and perform size normalization and pixel normalization processing on the aerial images of highways to obtain an input image. The second module is used to input the input image into a lightweight feature extraction network to extract multi-scale features, including shallow features, mid-level features, high-level features and deep features; The third module is used to input the deep features into the road strip context enhancement network, and to enhance the long strip continuous structure of the highway by using horizontal strip context, vertical strip context, local convolutional features and global semantic features to obtain road enhancement features; The fourth module is used to input the shallow features and the middle features into the boundary refinement network to generate a boundary attention map, and to perform weighted and residual enhancement on the initial boundary features based on the boundary attention map to obtain the boundary enhanced features; The fifth module is used to input the road enhancement features, high-level features, mid-level features, and boundary enhancement features into the multi-scale fusion decoder to obtain the decoded features; The sixth module is used to perform classification convolution and upsampling processing on the decoded features to obtain the semantic segmentation results of the highway image.

8. The semantic segmentation system for highway images according to claim 7, characterized in that, The third module includes: The first unit is used to compress the number of channels in deep features through 1×1 convolution; The second unit is used to extract horizontal strip context, vertical strip context, local convolution features and global semantic features from the compressed deep features through four parallel branches: horizontal average pooling, vertical average pooling, local convolution and global average pooling. The third unit is used to concatenate the outputs of the four branches along the channel dimension and fuse them through a 1×1 convolution to obtain the fused features; The fourth unit is used to add the fused features to the input deep features through residual connections to obtain road enhancement features.

9. The semantic segmentation system for highway images according to claim 7, characterized in that, The fourth module includes: The fifth unit is used to upsample the mid-level features to the same spatial scale as the shallow-level features and perform channel alignment through 1×1 convolution; The sixth unit is used to concatenate the middle-layer features with the shallow-layer features processed by 3×3 convolution, and then extract the initial boundary features through 3×3 convolution. The seventh unit is used to perform 1×1 convolution and Sigmoid activation on the initial boundary features to generate a boundary attention map.

10. The semantic segmentation system for highway images according to claim 7, characterized in that, The fifth module includes: The eighth unit is used to upsample the road enhancement features and fuse them with high-level features to obtain the first-level decoded features; The ninth unit is used to further upsample the first-level decoding features and fuse them with the mid-level features to obtain the second-level decoding features; The tenth unit is used to fuse the second-level decoding features with the boundary enhancement features to obtain the third-level decoding features.

Images