A remote sensing image road edge detection method and system

Abstract

The application provides a remote sensing image road edge detection method and system, and the method comprises the following steps: extracting a predicted candidate vertex set Q from a road image; selecting a candidate vertex from the predicted candidate vertex set Q as a predicted starting vertex v i *, generating a local feature input agent model thereof, outputting a next predicted vertex v i+1 *, and updating a generated road edge graph g i+1 * based on the next predicted vertex v t *, continuing prediction until the road edge graph g t stops growing; and traversing each candidate vertex in the predicted candidate vertex set Q, performing prediction based on the agent model to obtain a road edge detection result. The method mainly adopts an end-to-end iterative graph generation method, can directly obtain linear object information, is less affected by segmentation performance, can solve the influence caused by partial occlusion, and can obtain vector data of the road edge without complicated post-processing operations.

Description

Technical Field

[0001] This invention relates to the field of crowdsourced high-precision map production, and more specifically, to a method and system for detecting road edges in remote sensing images. Background Technology

[0002] Road outer edges play a crucial role in crowdsourced high-precision maps, providing key information for road planning and traffic safety. Currently, the detection of such linear objects using aerial imagery primarily relies on image segmentation and post-processing methods. The overall detection performance of these methods is typically limited by segmentation performance, significantly affected by occlusion from buildings and trees, and the post-processing operations are quite cumbersome. Summary of the Invention

[0003] This invention addresses the technical problems existing in the prior art by providing a method and system for detecting road edges in remote sensing images.

[0004] According to a first aspect of the present invention, a method for detecting road edges in remote sensing images is provided, comprising:

[0005] Feature map F is extracted from the road image, and prediction candidate vertices are extracted from the feature map F based on two segmentation heads to form a prediction candidate vertex set Q.

[0006] Select one candidate vertex from the prediction candidate vertex set Q as the prediction starting vertex v. i * Based on the feature map F and the currently generated road edge map g t Generate the predicted starting vertex v i * is a local feature, where i is a positive integer greater than 1;

[0007] Predict the starting vertex v i The local features of * are input into the surrogate model, and the next predicted vertex v is output. i+1 * and based on the next predicted vertex v i+1 *Update road edge map g t Continuously make predictions until the road edge map g t Stop growing;

[0008] Iterate through each candidate vertex in the predicted candidate vertex set Q as the starting vertex for prediction, and perform prediction based on the surrogate model to obtain the corresponding road edge map g. t ;

[0009] All road edge maps g t The road edge detection results that constitute the road image.

[0010] According to a second aspect of the present invention, a remote sensing image road edge detection system is provided, comprising:

[0011] The extraction module is used to extract feature map F from road images, and extract prediction candidate vertices from the feature map F based on two segmentation heads to form a prediction candidate vertex set Q;

[0012] The generation module is used to select a candidate vertex from the prediction candidate vertex set Q as the prediction starting vertex v. i * Based on the feature map F and the currently generated road edge map g t Generate the predicted starting vertex v i * is a local feature, where i is a positive integer greater than 1;

[0013] The prediction module is used to predict the starting vertex v. i The local features of * are input into the surrogate model, and the next predicted vertex v is output. i+1 * and based on the next predicted vertex v i+1 *Update road edge map g t Continuously make predictions until the road edge map g t Stop growth; traverse each candidate vertex in the predicted candidate vertex set Q as the starting vertex for prediction, and perform prediction based on the surrogate model to obtain the corresponding road edge map g. t All road edge maps g t The road edge detection results that constitute the road image.

[0014] The present invention provides a method and system for detecting road edges in remote sensing images. It mainly adopts an end-to-end iterative graph generation method, which can directly obtain linear object information. This method is less affected by segmentation performance, can solve the impact of partial occlusion, and can obtain vector data of road edges without cumbersome post-processing operations. Attached Figure Description

[0015] Figure 1 A flowchart of a remote sensing image road edge detection method provided by the present invention;

[0016] Figure 2 This is a schematic diagram of the overall method for road edge detection in remote sensing images.

[0017] Figure 3 This is a schematic diagram of the true vertex and the predicted vertex;

[0018] Figure 4 This is a schematic diagram of the structure of a remote sensing image road edge detection system provided by the present invention. Detailed Implementation

[0019] 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 embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. In addition, the technical features of the various embodiments or individual embodiments provided by the present invention can be arbitrarily combined with each other to form feasible technical solutions. Such combinations are not constrained by the order of steps and / or structural composition patterns, but must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0020] Based on the deficiencies in the background technology, this invention provides a method for detecting road edges in remote sensing images. This method breaks down the problem into two sub-problems: (1) finding the initial vertices of the road edges from the remote sensing images to start iterative graph growth; (2) iteratively generating road edge instances from the initial vertices. For problem (1), an algorithm is designed to generate an initial vertex candidate set based on the road edge segmentation results and the initial vertex heatmap. For problem (2), a new imitation learning algorithm is proposed. This algorithm dynamically generates training data to be labeled (i.e., expert demonstrations) based on the ground truth of the road edges. To solve the problem of learners failing to recover when anomalies occur due to a lack of learning, this algorithm also uses a primitive algorithm called Dataset (DAgger) to collect the learning behavior into the dataset to cover a larger data space to handle anomalies. The output of this algorithm is a graph G = (V, E) representing the road edges, and the vertex set V is generated by iteratively generating vertices vt = {x t s t Composed of}, where x t Represents two-dimensional coordinates, s t It is a variable that controls the stopping action; when s t When the value becomes 1, the iterative graph stops growing at the current vertex, and E is generated by connecting adjacent vertices during the graph's growth process.

[0021] See Figure 1 and Figure 2 The remote sensing image road edge detection method provided by this invention mainly includes:

[0022] S1, extract feature map F from road image, and extract prediction candidate vertices from feature map F based on two segmentation heads to form prediction candidate vertex set Q.

[0023] Understandably, satellite-captured road remote sensing images can be used, and feature maps F can be extracted from the input road remote sensing images using a feature pyramid (FPN).

[0024] Two segmentation heads are used to locate initial candidate vertices from feature map F. The first segmentation head obtains a binary segmentation result S of the road edges from feature map F, which includes multiple segmented road edges. These road edges are skeletonized, and shorter skeleton lines are filtered out. From the two endpoints of each remaining skeleton line, any one endpoint is selected as an initial candidate vertex. All initial candidate vertices constitute the first initial candidate vertex set. The second segmentation head predicts the heatmap H values ​​of the initial candidate vertices from feature map F. The heatmap H values ​​represent the probability distribution of the initial candidate vertices. Initial candidate vertices with heatmap H values ​​greater than a set threshold are selected to form the second initial candidate vertex set. The second segmentation head also predicts the heatmap H value of each initial candidate vertex in the first initial candidate vertex set, discarding those with heatmap H values ​​less than the set threshold. The first and second initial candidate vertex sets are merged to obtain the predicted candidate vertex set Q.

[0025] S2, Select a candidate vertex from the predicted candidate vertex set Q as the starting vertex v i Based on the feature map F and the currently generated road edge map g t Generate the starting vertex v i The local features of i, where i is a positive integer greater than 1.

[0026] Understandably, for each candidate vertex in the predicted candidate vertex set Q, it can be used as a starting vertex, and a corresponding road edge graph g can be generated based on the starting vertex. t Specifically, a candidate vertex is selected from the prediction candidate vertex set Q as the prediction starting vertex v. i * Combine the feature map F with the currently generated road edge map g t Connect to the new multi-channel feature map F t In the middle, here F t This can be viewed as imitating the environment in learning, with v i Centered on *, a square matrix block of a certain size can be used to map the multi-channel feature map F. t By cropping, the center is located at v. i *Feature map representation of v i * Local feature Crop.

[0027] S3 will predict the starting vertex v i The local features of * are input into the surrogate model, and the next predicted vertex v is output. i+1 * and based on the next predicted vertex v i+1*Update road edge map g t Continuously make predictions until the road edge map g t It stopped growing.

[0028] Understandably, the ordered true vertices obtained from the actual road on the ground are just some shape points that make up the road edge line. The road edge outline cannot be drawn using these points; the road edge outline is a vector graphic.

[0029] The initial candidate vertices determined in the initial candidate vertex set Q are only some shape points on the road edge line, which is not comprehensive. Therefore, the candidate vertices in the initial candidate vertex set Q are used as the starting vertices, and the corresponding road edge map is generated based on the surrogate model.

[0030] This invention can predict the starting vertex v i The local features are used to predict the next vertex based on the surrogate model. The training process of the surrogate model includes: obtaining an ordered sequence of actual vertices [v1, v2, ..., v] from the edges of the real road on the ground. i ..., v n ]; Obtain the predicted candidate vertex set from the corresponding road image, and select v from the predicted candidate vertex set. i * As the starting vertex for prediction, generate the starting vertex v for prediction. i * Local features; predict the starting vertex v i In the local feature input surrogate model, the output predicts the starting vertex v. i *The next predicted vertex v i+1 *; From the ordered actual vertex sequence [v1, v2, ..., v] of the real road edge on the ground. n Determining and predicting vertex v in [the context of the problem] i+1 *Corresponding actual vertex v i+1 And based on the actual vertex v i+1 And predicting vertex v i+1 The relationship between * determines the predicted vertex v i+1 *Whether to include the road edge map g t In the middle, based on the road edge map g t Perform an update; based on the updated road edge map g t Generate predicted vertex v i+1 The local features of * are input into the surrogate model, and the next predicted vertex v is output. i+2 *, ..., and so on, until the road edge graph g t No further growth; traverse each predicted vertex in the predicted candidate vertex set, generating all predicted vertices and their corresponding road edge graphs g. t Among them, each truth vertex v on the road i and the corresponding predicted vertex vi *See diagram for reference. Figure 3 .

[0031] Wherein, the ordered truth vertex sequence [v1, v2, ..., v] from the edge of the real road on the ground n Determining and predicting vertex v in [the context of the problem] i+1 *Corresponding truth vertex v i+1 This includes: an ordered sequence of truth vertices [v1, v2, ..., v] from the edges of real roads on the ground. n The vertex v in the truth value of ] i Among all subsequent truth vertices, find the vertex v that is the truth vertex. i The true vertex whose distance is greater than the first set distance threshold and whose distance is the smallest is the vertex with the predicted vertex v. i+1 *Corresponding truth vertex v i+1 Among them, the truth vertex v i To predict vertex v i * The corresponding truth vertex. Where, v i For the true vertex v i+1 The previous truth vertex, v i+1 *For predicting vertex v i * is the previous predicted vertex.

[0032] The search for the true vertex v i The true vertex whose distance is greater than the first set distance threshold and whose distance is the smallest is the vertex with the predicted vertex v. i+1 *Corresponding truth vertex v i+1 This includes: calculating the truth vertex v i-1 With the true vertex v i The direction vector between them is converted into a radian value r. i Obtain the ordered truth vertex sequence [v1, v2, ..., v] of the actual road edge on the ground. n The vertex v in the truth value of ] i For all subsequent truth vertices, calculate the radian value of the direction vector between every two adjacent truth vertices, and select the radian value that matches the radian value r. i The vertex whose difference is greater than the set radian threshold and whose minimum radian value corresponds to the starting true vertex v is used as the vertex with the predicted vertex v. i+1 *Corresponding truth vertex v i+1 .

[0033] Among them, the vertex based on truth value v i+1 And predicting vertex v i+1 The relationship between * determines the predicted vertex v i+1 *Whether to include the road edge map g t In the middle, based on the road edge map g tUpdates are performed, including: when predicting vertex v i+1 *and the corresponding truth vertex v i+1 If the distance between them is less than the second set distance threshold, then the prediction of vertex v is used. i+1 *For the already generated road edge map g t Update; and based on all predicted vertices v i+1 *For the already generated road edge map g t Update.

[0034] Understandably, if the predicted vertex v is output through the surrogate model... i+1 *and the corresponding truth vertex v i+1 If the distance between them is large, it indicates that the predicted value output by the surrogate model is not accurate enough. In this case, the predicted vertex v should be discarded. i+1 * and adjust the parameters of the surrogate model, then predict the next vertex based on the adjusted surrogate model. If the predicted vertex v i+1 *and the corresponding truth vertex v i+1 If the distance between them is relatively small, it indicates that the predicted vertex v output by the surrogate model is accurate. i+1 If the prediction is relatively accurate, then the predicted vertex v will be... i+1 *Enter the generated road edge map g t In the process, the generated road edge map g... t Perform an update based on the updated road edge map g. t Calculate the local features of the next predicted vertex.

[0035] It should be noted that, in order to ensure the consistency of data distribution during the training and testing phases of the surrogate model, it is also necessary to use free exploration to compensate for the shortcomings of restricted exploration, that is, directly predicting the vertex v. i+1 *Update road edge map g t No threshold condition is required.

[0036] The surrogate model is trained using the method described above, resulting in a trained surrogate model. By inputting the predicted starting vertex into the trained surrogate model, subsequent predicted vertices can be output, thereby generating the corresponding road edge map.

[0037] S4, traverse each candidate vertex in the predicted candidate vertex set Q as a starting vertex, and perform prediction based on the surrogate model to obtain the corresponding road edge map g. t All road edge maps g t The road edge detection results that constitute the road image.

[0038] Understandably, for each candidate starting vertex in the predicted candidate vertex set Q, a corresponding road edge map can be generated based on the surrogate model. However, these generated road edge maps may contain duplicate edge maps. Therefore, when traversing each candidate vertex in the predicted candidate vertex set Q, it is determined whether the candidate vertex in the predicted candidate vertex set Q exists in the already generated road edge map gt. If it exists, the candidate vertex is not processed; if it does not exist, the next predicted vertex and the generated road edge map are output based on the surrogate model. After traversing all the initial candidate vertices in the predicted candidate vertex set Q, multiple road edge maps are finally generated, which are the detection results of road edges in the remote sensing image.

[0039] It should be noted that, in the process of predicting the road edge map based on the initial candidate vertices in the predicted candidate vertex set Q, the predicted vertex v i+1 *Includes a coordinate header and a stop header to predict the x-axis of the next vertex. t+1 and s t+1 x t+1 Indicates the prediction of vertex v i+1 * Coordinates, s t+1 Indicates the prediction of vertex v i+1 The asterisk (*) indicates whether the prediction has stopped upon reaching this vertex. In other words, it indicates whether the road edge map has stopped growing upon reaching this vertex. If it stops growing, the prediction stops; otherwise, the prediction continues for the next vertex.

[0040] See Figure 4 This invention provides a remote sensing image road edge detection system, which includes an extraction module 401, a generation module 402, and a prediction module 403, wherein:

[0041] Extraction module 401 extracts feature map F from road imagery, and extracts prediction candidate vertices from feature map F based on two segmentation heads to form prediction candidate vertex set Q;

[0042] Generation module 402 is used to select a candidate vertex from the prediction candidate vertex set Q as the prediction starting vertex v. i * Based on the feature map F and the currently generated road edge map g t Generate the predicted starting vertex v i * is a local feature, where i is a positive integer greater than 1;

[0043] Prediction module 403 is used to predict the starting vertex v i The local features of * are input into the surrogate model, and the next predicted vertex v is output. i+1 * and based on the next predicted vertex v i+1 *Update road edge map g tContinuously make predictions until the road edge map g t Stop growth; traverse each candidate vertex in the predicted candidate vertex set Q as the starting vertex for prediction, and perform prediction based on the surrogate model to obtain the corresponding road edge map g. t All road edge maps g t The road edge detection results that constitute the road image.

[0044] It is understood that the remote sensing image road edge detection system provided by the present invention corresponds to the remote sensing image road edge detection method provided in the foregoing embodiments. The relevant technical features of the remote sensing image road edge detection system can be referred to the relevant technical features of the remote sensing image road edge detection method, and will not be repeated here.

[0045] This invention provides a method and system for detecting road edges in remote sensing images. The method mainly adopts an end-to-end iterative graph generation method, which can directly obtain linear object information. This method is less affected by segmentation performance, can solve the impact of partial occlusion, and can obtain vector data of road edges without cumbersome post-processing operations.

[0046] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0047] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0048] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0049] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0050] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0051] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0052] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for detecting road edges in remote sensing images, characterized in that, include: Feature map F is extracted from the road image, and prediction candidate vertices are extracted from the feature map F based on two segmentation heads to form a prediction candidate vertex set Q. Select one candidate vertex from the prediction candidate vertex set Q as the prediction starting vertex v. i * Based on the feature map F and the currently generated road edge map g t Generate the predicted starting vertex v i * is a local feature, where i is a positive integer greater than 1; Predict the starting vertex v i The local features of * are input into the surrogate model, and the next predicted vertex v is output. i+1 * and based on the next predicted vertex v i+1 *Update road edge map g t Continuously make predictions until the road edge map g t Stop growing; Iterate through each candidate vertex in the predicted candidate vertex set Q as the starting vertex for prediction, and perform prediction based on the surrogate model to obtain the corresponding road edge map g. t ; All road edge maps g t Road edge detection results constituting the road image; The step of extracting predicted candidate vertices from the feature map F based on two segmentation heads to form a predicted candidate vertex set Q includes: Based on the first segmentation head, the binary segmentation result S of the road edge line is obtained from the feature map F. The binary segmentation result S of the road edge line is skeletonized and the shorter skeleton lines are filtered out. From the two endpoints of each remaining skeleton line, any one endpoint is selected as the initial candidate vertex of the skeleton line. All the initial candidate vertices constitute the first initial candidate vertex set. Based on the second segmentation head, the heatmap H value of the initial candidate vertices is predicted from the feature map F. The heatmap H value represents the probability distribution of the initial candidate vertices. Initial candidate vertices with heatmap H values ​​greater than a set threshold are selected to form a second initial candidate vertex set. The heatmap H value of each first initial candidate vertex in the first initial candidate vertex set is predicted, and the first initial candidate vertices with heatmap H values ​​less than a set threshold are removed. The first initial candidate vertex set and the second initial candidate vertex set are merged to obtain the predicted candidate vertex set Q.

2. The road edge detection method according to claim 1, characterized in that, The feature map F and the currently generated road edge map g t Generate the predicted starting vertex v i Local features of * include: Combine the feature map F and the currently generated road edge map g t Connected to multi-channel feature map F t In the middle, for the current predicted starting vertex v i * Based on square matrix blocks of a set size, multi-channel feature maps F t Perform cropping to obtain a shape centered at v. i *Feature chart v i * Local feature Crop.

3. The road edge detection method according to claim 1, characterized in that, It also includes training agent models, including: Obtain the ordered truth vertex sequence [v1, v2, ..., v] of the actual road edge on the ground. i ..., v n ]; Obtain the predicted candidate vertex set from the corresponding road image, and select v from the predicted candidate vertex set. i * As the starting vertex for prediction, generate the starting vertex v for prediction. i * Local features; The predicted starting vertex v i In the local feature input surrogate model, the output predicts the starting vertex v. i *The next predicted vertex v i+1 *; From the ordered sequence of truth vertices [v1, v2, ..., v] at the edge of the real road on the ground. n Determining and predicting vertex v in [the context of the problem] i+1 *Corresponding truth vertex v i+1 And based on the truth vertex v i+1 And predicting vertex v i+1 The relationship between * determines the predicted vertex v i+1 *Whether to include the road edge map g t In the middle, based on the road edge map g t Update; Based on the updated road edge map g t Generate predicted vertex v i+1 The local features of * are input into the surrogate model, and the next predicted vertex v is output. i+2 *, ..., and so on, until the road edge graph g t No more growth; Iterate through each predicted vertex in the candidate vertex set to generate all predicted vertices and their corresponding road edge graphs g. t .

4. The road edge detection method according to claim 3, characterized in that, The ordered truth vertex sequence [v1, v2, ..., v] from the edge of the real road on the ground n Determining and predicting vertex v in [the context of the problem] i+1 *Corresponding truth vertex v i+1 ,include: From the ordered sequence of truth vertices [v1, v2, ..., v] at the edge of the real road on the ground. n The vertex v in the truth value of ] i Among all subsequent truth vertices, find the vertex v that is the truth vertex. i The true vertex whose distance is greater than the first set distance threshold and whose distance is the smallest is the vertex with the predicted vertex v. i+1 *Corresponding truth vertex v i+1 Among them, the truth vertex v i To predict vertex v i *The corresponding truth vertex.

5. The road edge detection method according to claim 4, characterized in that, The search for the true vertex v i The true vertex whose distance is greater than the first set distance threshold and whose distance is the smallest is the vertex with the predicted vertex v. i+1 *Corresponding truth vertex v i+1 ,include: Calculate the truth vertex v i-1 With the true vertex v i The direction vector between them is converted into a radian value r. i Obtain the ordered truth vertex sequence [v1, v2, ..., v] of the actual road edge on the ground. n The vertex v in the truth value of ] i For all subsequent truth vertices, calculate the radian value of the direction vector between every two adjacent truth vertices, and select the radian value that matches the radian value r. i The vertex whose difference is greater than the set radian threshold and whose minimum radian value corresponds to the starting true vertex v is used as the vertex with the predicted vertex v. i+1 *Corresponding truth vertex v i+1 .

6. The road edge detection method according to claim 3, characterized in that, The truth vertex v i+1 And predicting vertex v i+1 The relationship between * determines the predicted vertex v i+1 *Whether to include the road edge map g t In the middle, based on the road edge map g t Updates will be made, including: When predicting vertex v i+1 *and the corresponding truth vertex v i+1 If the distance between them is less than the second set distance threshold, then the prediction of vertex v is used. i+1 *For the already generated road edge map g t Update; And based on all predicted vertices v i+1 *For the already generated road edge map g t Update.

7. The road edge detection method according to claim 1, characterized in that, Each candidate vertex in the predicted candidate vertex set Q is used as a starting vertex, and prediction is performed based on the surrogate model to obtain the corresponding road edge map g. t ,include: Traverse any candidate vertex in the predicted candidate vertex set Q, and determine whether the candidate vertex is in the generated road edge graph g. t If a candidate vertex exists, then none of the candidate vertices will be processed. If it does not exist, the next predicted vertex and the generated road edge map are output based on the surrogate model for any candidate vertex.

8. A remote sensing image road edge detection system, characterized in that, include: The extraction module is used to extract feature map F from road images, and extract prediction candidate vertices from the feature map F based on two segmentation heads to form a prediction candidate vertex set Q; The production module is used to select a candidate vertex v from the prediction candidate vertex set Q as the prediction starting vertex. i * Based on the feature map F and the currently generated road edge map g t Generate the predicted starting vertex v i * is a local feature, where i is a positive integer greater than 1; The prediction module is used to predict the starting vertex v. i The local features of * are input into the surrogate model, and the next predicted vertex v is output. i+1 * and based on the next predicted vertex v i+1 *Update road edge map g t Continuously make predictions until the road edge map g t Stop growth; traverse each candidate vertex in the predicted candidate vertex set Q as the starting vertex for prediction, and perform prediction based on the surrogate model to obtain the corresponding road edge map g. t All road edge maps g t Road edge detection results constituting the road image; The step of extracting predicted candidate vertices from the feature map F based on two segmentation heads to form a predicted candidate vertex set Q includes: Based on the first segmentation head, the binary segmentation result S of the road edge line is obtained from the feature map F. The binary segmentation result S of the road edge line is skeletonized and the shorter skeleton lines are filtered out. From the two endpoints of each remaining skeleton line, any one endpoint is selected as the initial candidate vertex of the skeleton line. All the initial candidate vertices constitute the first initial candidate vertex set. Based on the second segmentation head, the heatmap H value of the initial candidate vertices is predicted from the feature map F. The heatmap H value represents the probability distribution of the initial candidate vertices. Initial candidate vertices with heatmap H values ​​greater than a set threshold are selected to form a second initial candidate vertex set. The heatmap H value of each first initial candidate vertex in the first initial candidate vertex set is predicted, and the first initial candidate vertices with heatmap H values ​​less than a set threshold are removed. The first initial candidate vertex set and the second initial candidate vertex set are merged to obtain the predicted candidate vertex set Q.

Images