A method and apparatus for extracting a road from a remote sensing image

Abstract

The present disclosure relates to a method and device for extracting roads from remote sensing images. A method for training a model for extracting roads from remote sensing images is provided, the method being performed in the cloud, the method comprising: annotating a remote sensing image to obtain a first image and a second image, wherein the first image marks out road pavements in the remote sensing image, and the second image marks out road edges in the remote sensing image; obtaining a third image from the remote sensing image based on an edge detection algorithm, wherein the third image marks out road edges in the remote sensing image; inputting the third image into an edge extraction network in the model and training the edge extraction network using the second image; and inputting the remote sensing image into a segmentation network in the model and training the segmentation network using the first image and edge features from the edge extraction network.

Description

Technical Field

[0001] This disclosure relates to the fields of image processing and artificial intelligence, and more specifically to a method and apparatus for extracting roads from remote sensing images. Background Technology

[0002] Road extraction from remote sensing images generally refers to the automated or semi-automated extraction of roads from these images. How to quickly and effectively extract roads from remote sensing images has always been a challenging and crucial research area. Traditional road data acquisition primarily relies on manual methods, such as real-time surveys and reconnaissance by field teams. While this method can obtain relatively accurate road information, it requires significant manpower and resources, has a slow update speed, and is also inherently dangerous. Deep learning can effectively overcome the shortcomings of traditional manual feature selection. With the continuous development of deep learning networks, the information that deep learning can extract is becoming increasingly rich, thus enabling its application to a variety of tasks.

[0003] On the other hand, as the amount of data required for processing continues to increase, relying solely on local data storage and model training can no longer meet user needs. However, cloud transmission leads to problems such as long transmission links, client response delays, and network jitter, resulting in poor fault tolerance.

[0004] Therefore, there is a need for methods and apparatuses that can overcome the aforementioned deficiencies in extracting roads from remote sensing images. Summary of the Invention

[0005] A brief overview of this disclosure is given below to provide a basic understanding of some aspects of it. However, it should be understood that this overview is not an exhaustive summary of this disclosure. It is not intended to identify key or essential parts of this disclosure, nor is it intended to limit the scope of this disclosure. Its purpose is merely to present certain concepts of this disclosure in a simplified form as a prelude to the more detailed description that follows.

[0006] According to a first aspect of this disclosure, a method is provided for training a model for extracting roads from remotely sensed images, the method being executed in the cloud, the method comprising: annotating the remotely sensed images to obtain a first image and a second image, wherein the first image marks road surfaces in the remotely sensed images, and the second image marks road edges in the remotely sensed images; obtaining a third image from the remotely sensed images based on an edge detection algorithm, wherein the third image marks road edges in the remotely sensed images; inputting the third image into an edge extraction network in the model and training the edge extraction network using the second image; and inputting the remotely sensed images into a segmentation network in the model and training the segmentation network using the first image and edge features from the edge extraction network.

[0007] According to a second aspect of this disclosure, a method for extracting roads from a remote sensing image is provided, the method being performed at an edge, the method comprising: sending a request to a cloud to obtain a trained model, the trained model being trained according to the method of the first aspect; receiving a request from a terminal for extracting roads from the remote sensing image, wherein the request includes the remote sensing image; performing calculations on the remote sensing image using the trained model to obtain a model output, the model output indicating roads in the remote sensing image; and returning the model output to the terminal.

[0008] According to a third aspect of this disclosure, a training apparatus is provided for training a model for extracting roads from remotely sensed images. The training apparatus is located in the cloud and is configured to: annotate the remotely sensed images to obtain a first image and a second image, wherein the first image marks road surfaces in the remotely sensed images, and the second image marks road edges in the remotely sensed images; obtain a third image from the remotely sensed images based on an edge detection algorithm, wherein the third image marks road edges in the remotely sensed images; input the third image into an edge extraction network in the model and train the edge extraction network using the second image; and input the remotely sensed images into a segmentation network in the model and train the segmentation network using the first image and edge features from the edge extraction network.

[0009] According to a fourth aspect of this disclosure, an apparatus for extracting roads from a remote sensing image is provided, the apparatus being located at an edge, the apparatus comprising: a decision module configured to: send a request to a cloud to obtain a trained model, the trained model being trained according to the method described in the first aspect of this disclosure; receive a request from a terminal for extracting roads from a remote sensing image, wherein the request includes the remote sensing image; and return a model output to the terminal, the model output indicating roads in the remote sensing image; and a calculation module configured to perform calculations on the remote sensing image using the trained model to obtain the model output.

[0010] According to a fifth aspect of this disclosure, a non-transitory computer-readable storage medium is provided, having a program stored thereon, characterized in that, when the program is executed by a computer, it causes the computer to perform the method according to the first aspect.

[0011] According to a sixth aspect of this disclosure, a non-transitory computer-readable storage medium is provided, having a program stored thereon, characterized in that, when the program is executed by a computer, it causes the computer to perform the method according to the second aspect.

[0012] According to a seventh aspect of this disclosure, an apparatus is provided for training a model for extracting roads from remotely sensed images, comprising a memory and a processor, the memory being communicatively coupled to the processor, the memory storing a program that, when executed by the processor, causes the processor to perform the method according to the first aspect.

[0013] According to an eighth aspect of this disclosure, an apparatus for extracting roads from remotely sensed images is provided, comprising a memory and a processor, the memory being communicatively coupled to the processor, the memory storing a program that, when executed by the processor, causes the processor to perform the method according to the second aspect.

[0014] According to a ninth aspect of this disclosure, a computer program product is provided, comprising a computer program that, when executed by a processor, implements the method according to the first aspect.

[0015] According to a tenth aspect of this disclosure, a computer program product is provided, comprising a computer program that, when executed by a processor, implements the method described according to a second aspect.

[0016] The methods and apparatus proposed in this disclosure can effectively achieve high bandwidth and low latency, fully utilize the advantages of near-end deployment, and improve the user experience.

[0017] Other features and advantages of this disclosure will become clearer from the following detailed description of exemplary embodiments with reference to the accompanying drawings. Attached Figure Description

[0018] The accompanying drawings, which form part of this specification, illustrate embodiments of this disclosure and, together with the specification, serve to explain the principles of this disclosure.

[0019] This disclosure will be more clearly understood with reference to the accompanying drawings and the following detailed description, in which:

[0020] Figure 1 The overall architecture of a system for extracting roads from remotely sensed images according to embodiments of the present disclosure is shown;

[0021] Figure 2 A schematic diagram of a cloud-edge collaborative architecture according to an embodiment of the present disclosure is shown;

[0022] Figure 3 A schematic diagram of the architecture of a neural network according to an embodiment of the present disclosure is shown;

[0023] Figure 4 A schematic diagram of a neural network according to an embodiment of the present disclosure is shown;

[0024] Figure 5 A flowchart illustrating a method for training a model for extracting roads from remotely sensed images according to embodiments of the present disclosure is shown;

[0025] Figure 6 An exemplary configuration of a computing device that can implement embodiments of the present disclosure is shown. Detailed Implementation

[0026] The following detailed description is based on the accompanying drawings and provides various exemplary embodiments of the present disclosure to aid in a comprehensive understanding. Various details are included in the following description to aid understanding; however, these details are considered exemplary only and not intended to limit the present disclosure, which is defined by the appended claims and their equivalents. The words and phrases used in the following description are intended only to provide a clear and consistent understanding of the present disclosure. Additionally, descriptions of well-known structures, functions, and configurations may have been omitted for clarity and brevity. Those skilled in the art will recognize that various changes and modifications can be made to the examples described herein without departing from the spirit and scope of the present disclosure.

[0027] Figure 1 The overall architecture of a system for extracting roads from remotely sensed images according to embodiments of the present disclosure is shown. Figure 1As shown, the system can include three parts: cloud, edge, and terminal. The edge can include multiple edge nodes. The cloud can be configured to perform data storage, AI training (i.e., model training), and control decision-making. Edge nodes can be configured to perform storage, computation, and decision-making. Terminals can be configured to perform remote sensing image acquisition and control decision-making.

[0028] Figure 2 A schematic diagram of a cloud-edge collaborative architecture according to an embodiment of the present disclosure is shown. (As follows) Figure 2 As shown, the cloud-edge collaborative architecture can include a cloud 210, an MEC (Multi-Access Edge Computing) 220, and an edge device 230.

[0029] The cloud 210 can be configured to receive data from the MEC 220, use the received data for training, and distribute the trained model. The cloud 210 may include a data storage module 211 and a model training module 212. The data storage module 211 can be configured to store remote sensing image data from the MEC 220 and the model trained by the model training module 212. The model training module 212 can be configured to perform data processing and model training based on the remote sensing image data from the MEC 220. In embodiments according to this disclosure, the cloud 210 may also include a decision module, configured, for example, to determine whether to request the synchronization of remote sensing image data stored at the edge to the cloud, control the distribution of the model, and cloud-edge collaboration.

[0030] MEC 220 can be configured to perform data collaboration, etc. For example, MEC 220 can receive data uploaded from edge device 230 and send data uploaded from edge device 230 to cloud 210.

[0031] Edge device 230 can be configured to collect remote sensing image data from the terminal, upload the collected data, perform calculations based on a distributed model, and feed the calculated results back to the terminal. Edge device 230 may include a data acquisition module 231 and a calculation module 232. Data acquisition module 231 can be configured to: receive remote sensing image data from the terminal and upload the remote sensing image data to MEC 220; send the remote sensing image data to be identified to the calculation module 232 and receive the calculation results from the calculation module 232; and feed the calculation results back to the terminal. Calculation module 232 can be configured to quickly perform calculations on the remote sensing image data to be identified based on the distributed model to extract roads from the remote sensing image data to be identified. In embodiments according to this disclosure, calculation module 232 may also include a storage module for storing temporary data and models distributed from the cloud.

[0032] In embodiments according to this disclosure, the terminal may include a request module and a receiving module. The request module may be configured to send a request to the edge terminal, and the receiving module may be configured to receive the calculated results from the edge terminal and display the results, etc.

[0033] The method for extracting roads from remote sensing image data to be identified based on cloud-edge collaboration according to this disclosure may include the following steps:

[0034] Step 1: Train the model in the cloud and distribute the model to the edge nodes via MEC based on the cloud's decision.

[0035] Step 2: The end user sends a request to send the remote sensing image data to be identified to the edge node. The edge node temporarily stores the remote sensing image data to be identified in its storage module through data acquisition. The edge node performs calculations on the remote sensing image data to be identified based on the model and returns the final calculation result to the end user.

[0036] Step 3: Based on the cloud's decision, determine whether to synchronize the remote sensing image data in the edge storage module to the cloud. In embodiments according to this disclosure, after synchronizing the image data to the cloud, the image data stored locally in the edge storage module can be deleted. In embodiments according to this disclosure, the edge node can determine a security level based on the cloud's decision to choose whether to periodically synchronize the remote sensing image data to the cloud.

[0037] Step 4: Once a sufficient amount of new image data has been collected in the cloud, fine-tune the model and distribute the trained model to the edge nodes for model updates.

[0038] Step 5: In the event that the model is accidentally deleted at the edge, request the cloud to redeploy the model.

[0039] Figure 3 and Figure 4 A schematic diagram of the architecture of a neural network (i.e., a model for extracting roads from remotely sensed images) according to an embodiment of the present disclosure is shown.

[0040] like Figure 3 and Figure 4 As shown, the backbone networks of the neural network are the edge extraction network (Edgenet) and the segmentation network (with Efficientnet as the backbone).

[0041] The method for training a neural network according to this disclosure may include the following steps:

[0042] Step 1: Process the remote sensing image data I sThe road areas are labeled, and the road labeling results are represented as image I. b and Image I e Image I b Mark the road surface in the remote sensing image, where the road surface area is denoted as R, and the non-road surface area (i.e., the background area) is denoted as B. Image I e The edges of roads in the remotely sensed image are marked, wherein the edge regions on both sides of the road are denoted as G. In embodiments according to this disclosure, the marking can be done manually.

[0043] For example, it is possible to analyze remote sensing image I s Manual annotation is performed to obtain road image I. b With edge image I e Image I b This includes the road surface area R (pixel value (255, 0, 0)) and the background area B (pixel value (0, 0, 0)); I e This includes the edge region G (pixel value (255, 0, 0)). For example, I can be... b and I e After conversion, the pixel values ​​of R and G are 1, and the pixel value of B is 0.

[0044] Step 2: Extract edges from the remote sensing image using an edge detection algorithm to obtain image I. c Image I c The road edges in the remote sensing image are marked.

[0045] In embodiments according to this disclosure, the edge detection algorithm may be the Canny algorithm.

[0046] Step 3: Transfer image I s I b I e I c Perform cropping and assemble corresponding input sample pairs (I m , L m C m E m In embodiments according to this disclosure, an image I of size W x H can be used. s I b I e I c Perform batch cropping to crop it into several images of size wxh, and form corresponding input sample pairs.

[0047] For example, GDAL can be used to target I s I b I e I cPerform batch cropping, with images sized to 512x512, to form input sample pairs (I m , L m C m E m ).

[0048] Step 4: Train the neural network, whose backbone networks are an edge extraction network and a segmentation network.

[0049] Specifically, training a neural network may include: selecting an optimization function to perform backpropagation on the network, while setting hyperparameters such as the number of iterations and the learning rate; and inputting sample pairs (I... m , L m C m E m Input the network for training; save the trained model after the network converges.

[0050] Edgenet's input includes C m Edge features are extracted using a relatively shallow network, and based on the edge features and E... m Calculate the loss function loss1.

[0051] The input to the segmentation network includes I m Features are extracted using the Efficient algorithm and upsampled through several iterations of Upsampling and Conv to obtain the final Feature_map. Furthermore, the Feature_map and Edge_features are concatenated, and the Activate operation is used to obtain the neural network output. The loss function loss2 is calculated based on the output and L. m .

[0052] For example, the edge extraction network could be Edgenet, which can select the fourth convolutional layer of VGG16 as the backbone, with an output of (256, 256, 128), and add a Deconv layer and a Sigmoid function to obtain the edge feature `Edge_feature`, with a size of (512, 512, 1). Based on the edge feature `Edge_feature` and E... m Calculate the loss function loss1, which is preferably bce_loss.

[0053] For example, the segmentation network can use EfficientNet as the backbone, preferably EfficientNetB5, with an output size of (16, 16, 2048). This output size is then restored to (512, 512, 2048) through a 32x upsampling process and passed through a Conv layer to obtain a feature map with a size of (512, 512, 1).

[0054] For example, the feature_map is concatenated with the edge feature_feature, and then processed by the sigmoid activation function, resulting in an output of size (512, 512, 1). Based on the output and L... m Calculate the loss function loss2, which is preferably dice_loss or focal_loss.

[0055] For example, the optimization function is RMSprop.

[0056] Figure 5 A flowchart illustrating a method for training a model to extract roads from remotely sensed images according to embodiments of the present disclosure is shown. This method can be performed in the cloud.

[0057] like Figure 5 As shown, at point S501, the remote sensing image I can be... s Annotation is performed to obtain the first image I. b Second image I e Among them, the first image I b The road surface in the remote sensing image is marked in the second image I. e The edges of roads in the remote sensing image are marked.

[0058] At point S502, edge detection algorithms can be used to extract data from remote sensing image I. s Obtain the third image I c Among them, the third image I c The road edges in the remote sensing image are marked.

[0059] At S503, the third image I can be... c Input the edge extraction network in the model and utilize the second image I e (Directly using E) m E m For the second image I e The edge extraction network is trained by processing the data.

[0060] At S504, the remote sensing image I can be... sInput (directly put I) m Input, I m For remote sensing image I s The segmentation network in the model obtained after processing is used, and the first image I is used. b The segmentation network is trained using edge features from the edge extraction network.

[0061] Figure 6 An exemplary configuration of a computing device 600 capable of implementing embodiments of the present disclosure is shown.

[0062] Computing device 600 is an example of a hardware device capable of applying the foregoing aspects of this disclosure. Computing device 600 can be any machine configured to perform processing and / or computation. Computing device 600 can be, but is not limited to, a workstation, server, desktop computer, laptop computer, tablet computer, personal data assistant (PDA), smartphone, in-vehicle computer, or a combination thereof.

[0063] like Figure 6 As shown, computing device 600 may include one or more components that can be connected to or communicate with bus 602 via one or more interfaces. Bus 602 may include, but is not limited to, Industry Standard Architecture (ISA) bus, Micro Channel Architecture (MCA) bus, Enhanced ISA (EISA) bus, Video Electronics Standards Association (VESA) local bus, and Peripheral Component Interconnect (PCI) bus. Computing device 600 may include, for example, one or more processors 604. The one or more processors 604 may be any kind of processor and may include, but is not limited to, one or more general-purpose processors or special-purpose processors (such as dedicated processing chips). The processor may, for example, be configured to implement the methods described above for training a model for extracting roads from remote sensing images and for extracting roads from remote sensing images.

[0064] The computing device 600 may also include or be connected to a non-transitory storage device 614, which may be any non-transitory storage device capable of storing data, and may include, but is not limited to, disk drives, optical storage devices, solid-state storage, floppy disks, flexible disks, hard disks, magnetic tapes or any other magnetic media, compressed disks or any other optical media, cache memory and / or any other storage chip or module, and / or any other medium from which a computer may read data, instructions and / or code. The computing device 600 may also include random access memory (RAM) 610 and read-only memory (ROM) 612. ROM 612 may store executable programs, utilities, or processes in a non-volatile manner. RAM 610 provides volatile data storage and stores instructions related to the operation of the computing device 600.

[0065] In summary, according to a first aspect of this disclosure, a method for training a model to extract roads from remotely sensed images is provided, the method being executed in the cloud, the method comprising: annotating the remotely sensed images to obtain a first image and a second image, wherein the first image marks road surfaces in the remotely sensed images, and the second image marks road edges in the remotely sensed images; obtaining a third image from the remotely sensed images based on an edge detection algorithm, wherein the third image marks road edges in the remotely sensed images; inputting the third image into an edge extraction network in the model and training the edge extraction network using the second image; and inputting the remotely sensed images into a segmentation network in the model and training the segmentation network using the first image and edge features from the edge extraction network.

[0066] In an embodiment according to this disclosure, the model is configured to combine edge features from the edge extraction network with segmentation features from the segmentation network to obtain a model output indicating a road in the remote sensing image.

[0067] In embodiments according to this disclosure, the loss function of the edge extraction network is calculated based on the edge features and the second image, and the loss function of the segmentation network is calculated based on the model output and the first image.

[0068] According to a second aspect of this disclosure, a method for extracting roads from a remote sensing image is provided, the method being performed at an edge, the method comprising: sending a request to a cloud to obtain a trained model, the trained model being trained according to the method of the first aspect; receiving a request from a terminal for extracting roads from the remote sensing image, wherein the request includes the remote sensing image; performing calculations on the remote sensing image using the trained model to obtain a model output, the model output indicating roads in the remote sensing image; and returning the model output to the terminal.

[0069] In embodiments according to this disclosure, the method further includes: synchronizing remote sensing images received from the terminal to the cloud for use in training or updating the model.

[0070] In an embodiment according to this disclosure, the method further includes: deleting the remote sensing image locally from the edge in response to the remote sensing image being synchronized to the cloud.

[0071] In embodiments according to this disclosure, the method further includes: sending a request to the cloud to retrieve the model again if the model is updated in the cloud or if the model is deleted locally at the edge.

[0072] According to a third aspect of this disclosure, a training apparatus is provided for training a model for extracting roads from remotely sensed images. The training apparatus is located in the cloud and is configured to: annotate the remotely sensed images to obtain a first image and a second image, wherein the first image marks road surfaces in the remotely sensed images, and the second image marks road edges in the remotely sensed images; obtain a third image from the remotely sensed images based on an edge detection algorithm, wherein the third image marks road edges in the remotely sensed images; input the third image into an edge extraction network in the model and train the edge extraction network using the second image; and input the remotely sensed images into a segmentation network in the model and train the segmentation network using the first image and edge features from the edge extraction network.

[0073] In an embodiment according to this disclosure, the model is configured to combine edge features from the edge extraction network with segmentation features from the segmentation network to obtain a model output indicating a road in the remote sensing image.

[0074] In embodiments according to this disclosure, the loss function of the edge extraction network is calculated based on the edge features and the second image, and the loss function of the segmentation network is calculated based on the model output and the first image.

[0075] According to a fourth aspect of this disclosure, an apparatus for extracting roads from a remote sensing image is provided, the apparatus being located at an edge, the apparatus comprising: a decision module configured to: send a request to a cloud to obtain a trained model, the trained model being trained according to the method described in the first aspect of this disclosure; receive a request from a terminal for extracting roads from a remote sensing image, wherein the request includes the remote sensing image; and return a model output to the terminal, the model output indicating roads in the remote sensing image; and a calculation module configured to perform calculations on the remote sensing image using the trained model to obtain the model output.

[0076] In embodiments according to this disclosure, the decision module is further configured to synchronize remote sensing images received from the terminal to the cloud for use in training or updating the model.

[0077] In an embodiment according to this disclosure, the decision module is further configured to: delete the remote sensing image locally from the edge in response to the remote sensing image being synchronized to the cloud.

[0078] In embodiments according to this disclosure, the decision module is further configured to: send a request to the cloud to re-acquire the model if the model is updated in the cloud or if the model is deleted locally at the edge.

[0079] According to a fifth aspect of this disclosure, a non-transitory computer-readable storage medium is provided, having a program stored thereon, characterized in that, when the program is executed by a computer, it causes the computer to perform the method according to the first aspect.

[0080] According to a sixth aspect of this disclosure, a non-transitory computer-readable storage medium is provided, having a program stored thereon, characterized in that, when the program is executed by a computer, it causes the computer to perform the method according to the second aspect.

[0081] According to a seventh aspect of this disclosure, an apparatus is provided for training a model for extracting roads from remotely sensed images, comprising a memory and a processor, the memory being communicatively coupled to the processor, the memory storing a program that, when executed by the processor, causes the processor to perform the method according to the first aspect.

[0082] According to an eighth aspect of this disclosure, an apparatus for extracting roads from remotely sensed images is provided, comprising a memory and a processor, the memory being communicatively coupled to the processor, the memory storing a program that, when executed by the processor, causes the processor to perform the method according to the second aspect.

[0083] According to a ninth aspect of this disclosure, a computer program product is provided, comprising a computer program that, when executed by a processor, implements the method according to the first aspect.

[0084] According to a tenth aspect of this disclosure, a computer program product is provided, comprising a computer program that, when executed by a processor, implements the method described according to a second aspect.

[0085] The method and apparatus disclosed herein effectively incorporate edge information and use it as a constraint, thereby improving the accuracy of road extraction. Unlike traditional private (local) deployments and cloud deployments, this cloud-edge collaboration effectively reduces transmission latency, saves computing resources, and improves user experience.

[0086] The subject matter of this disclosure is provided as examples of apparatus, systems, methods, and programs for performing the features described herein. However, other features or variations are contemplated in addition to those described above. It is anticipated that the components and functions of this disclosure can be implemented using any emerging techniques that may replace any of the above-described implementations.

[0087] Furthermore, the above description provides examples and does not limit the scope, applicability, or configuration set forth in the claims. Changes may be made to the function and arrangement of the elements discussed without departing from the spirit and scope of this disclosure. Various processes or components may be appropriately omitted, substituted, or added in various embodiments. For example, features described with respect to certain embodiments may be combined in other embodiments.

[0088] Furthermore, in the description of this disclosure, the terms “first,” “second,” “third,” etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or order.

[0089] Similarly, although the operations are depicted in a specific order in the accompanying drawings, this should not be construed as requiring the operations to be performed in the specific order shown or in sequential order, or requiring the execution of all illustrated operations to achieve the desired result. In some cases, multitasking and parallel processing can be advantageous.

Claims

1. A method for training a model to extract roads from remotely sensed images, the method being performed in the cloud, the method comprising: The remote sensing images are annotated to obtain a first image and a second image, wherein the road surface in the remote sensing image is marked in the first image and the road edge in the remote sensing image is marked in the second image; A third image is obtained from the remote sensing image based on an edge detection algorithm, wherein the road edges in the remote sensing image are marked in the third image, and the edge detection algorithm is the Canny algorithm; The third image is input into the edge extraction network in the model, and the edge extraction network is trained using the second image; The remote sensing image is input into the segmentation network of the model, and the segmentation network is trained using the first image and edge features from the edge extraction network; and Edge features from the edge extraction network are combined with segmentation features from the segmentation network to obtain a model output that indicates a road in the remote sensing image.

2. The method according to claim 1, wherein, The loss function of the edge extraction network is calculated based on the edge features and the second image, and the loss function of the segmentation network is calculated based on the model output and the first image.

3. A method for extracting roads from a remotely sensed image, the method being performed at the edge end, the method comprising: Send a request to the cloud to obtain a trained model, wherein the trained model is trained according to the method of claim 1 or 2; Receive a request from the terminal for extracting roads from a remote sensing image, wherein the request includes the remote sensing image; The remotely sensed image is processed using a trained model to obtain a model output indicating roads in the remotely sensed image; and The model output is returned to the terminal.

4. The method according to claim 3, wherein, The method further includes: The remote sensing images received from the terminal will be synchronized to the cloud for use in model training or updates.

5. The method according to claim 4, wherein, The method further includes: In response to the remote sensing image being synchronized to the cloud, the remote sensing image is deleted locally from the edge device.

6. The method according to claim 3, wherein, The method further includes: If the model is updated in the cloud or deleted locally at the edge, a request is sent to the cloud to retrieve the model again.

7. A training apparatus for training a model for extracting roads from remotely sensed images, the training apparatus being located in the cloud, the training apparatus being configured to: The remote sensing images are labeled to obtain a first image and a second image, wherein, The first image marks the road surface in the remote sensing image, and the second image marks the road edges in the remote sensing image; A third image is obtained from the remote sensing image based on an edge detection algorithm, wherein the road edges in the remote sensing image are marked in the third image, and the edge detection algorithm is the Canny algorithm; The third image is input into the edge extraction network in the model, and the edge extraction network is trained using the second image; The remote sensing image is input into the segmentation network of the model, and the segmentation network is trained using the first image and edge features from the edge extraction network; and Edge features from the edge extraction network are combined with segmentation features from the segmentation network to obtain a model output that indicates a road in the remote sensing image.

8. The training apparatus according to claim 7, wherein, The loss function of the edge extraction network is calculated based on the edge features and the second image, and the loss function of the segmentation network is calculated based on the model output and the first image.

9. An apparatus for extracting roads from a remotely sensed image, the apparatus being located at an edge, the apparatus comprising: The decision module is configured as follows: A request is sent to the cloud to obtain a trained model, wherein the trained model is obtained by the method according to claim 1 or 2. The terminal receives a request to extract roads from a remote sensing image, wherein the request includes the remote sensing image, and The model output is returned to the terminal, indicating a road in the remotely sensed image; and The computation module is configured to perform computations on the remotely sensed image using a trained model to obtain the model output.

10. The apparatus according to claim 9, wherein, The decision module is also configured to: The remote sensing images received from the terminal will be synchronized to the cloud for use in model training or updates.

11. The apparatus according to claim 10, wherein, The decision module is also configured to: In response to the remote sensing image being synchronized to the cloud, the remote sensing image is deleted locally from the edge device.

12. The apparatus according to claim 9, wherein, The decision module is also configured to: If the model is updated in the cloud or deleted locally at the edge, a request is sent to the cloud to retrieve the model again.

13. A non-transitory computer-readable storage medium having a program stored thereon, characterized in that, When the program is executed by a computer, it causes the computer to perform the method according to claim 1 or 2.

14. A non-transitory computer-readable storage medium having a program stored thereon, characterized in that, When the program is executed by a computer, it causes the computer to perform the method according to any one of claims 3-6.

15. An apparatus for training a model for extracting roads from remotely sensed images, comprising a memory and a processor, the memory being communicatively coupled to the processor, the memory storing a program that, when executed by the processor, causes the processor to perform the method according to claim 1 or 2.

16. An apparatus for extracting roads from remotely sensed images, comprising a memory and a processor, the memory being communicatively coupled to the processor, the memory storing a program that, when executed by the processor, causes the processor to perform the method according to any one of claims 3-6.

17. A computer program product comprising a computer program that, when executed by a processor, implements the method according to claim 1 or 2.

18. A computer program product comprising a computer program that, when executed by a processor, implements the method according to any one of claims 3-6.

Images