A remote sensing image vector contour extraction method, system, storage medium and device

Abstract

The application discloses a remote sensing image vector building contour extraction method, constructs a building vector contour extraction model, obtains a multi-dimensional feature map through a feature pyramid network, generates building position coordinates through a region proposal network, generates overall contour features, corner point features and edge features through a Transform network, the three features are constrained through a correlation loss function, point coordinates are obtained through the corner point features, the corner point features, the edge features and the point coordinates are combined to input an RNN module to obtain position features of a next node, the position features of the next node and Fourier shape descriptor features of the overall contour features are fused, and overall vector contour information is obtained through iteration. The application can improve the misjudgment phenomenon of a building ridge line in a remote sensing image, effectively constrain an incomplete building shape, reduce topological errors in a building contour, and reduce errors in a building edge point extraction process.

Description

Technical Field

[0001] This invention relates to the field of remote sensing application technology, and in particular to a method, system, storage medium, and device for extracting vector contours from remote sensing images. Background Technology

[0002] Using remote sensing satellite data for Earth observation is a rapid method for acquiring information about the Earth's surface. The vast resources of remote sensing data make large-scale, all-round, all-time, and multi-dimensional observation and monitoring possible. High-resolution remote sensing images contain a rich variety of target objects, with buildings being the most typical man-made features. High-resolution remote sensing images can capture rich scene details, and the color, texture, and structural topology of buildings can be better characterized. Because the visual perception and spectral bands of high-resolution remote sensing images are similar to human visual perception, they have better visibility and interpretability. Furthermore, their large scale, wide range, and abundant data provide new ideas for large-scale building outline extraction research. Fully and promptly utilizing high-resolution visible light remote sensing images for automatic and accurate building extraction has broad practical significance.

[0003] In recent years, with the continuous improvement of the resolution of acquired visible light remote sensing images, higher requirements have been placed on the accuracy of extracted building outlines. Traditional morphological algorithm-based building outline extraction methods are no longer used due to the difficulty in obtaining prior shape knowledge. Deep learning-based building outline extraction methods have evolved from semantic segmentation methods, instance segmentation methods, single-building vector outline extraction methods, to multi-building vector outline extraction methods. Semantic segmentation-based building outline extraction methods can accurately locate buildings in remote sensing images and extract relatively complete pixel-level building outlines through powerful image feature extraction capabilities. However, semantic segmentation cannot distinguish individual buildings in dense building scenes, which leads to reduced accuracy in remote sensing interpretation applications based on building outlines. Therefore, to accurately distinguish the location of individual building outlines in remote sensing scenes, instance segmentation algorithms have been introduced into building outline extraction research. Instance segmentation methods extract the bounding rectangle of the building outline and then perform instance segmentation on the targets within the rectangle, thereby achieving pixel-level building object extraction. However, most smart city construction research based on building outline locations relies on vector-based building outlines. Converting pixel-level building information extracted by instance segmentation methods to vector format introduces a series of problems (loss of spatial correlation information between key corner points during the conversion). Therefore, directly extracting vector-level building corner points from raster remote sensing imagery has become a current research focus. Methods for extracting single building vector outlines have also evolved from a practical application perspective, and current methods can now extract information from multiple buildings simultaneously. However, methods that extract key corner points and then sequentially connect them to form a complete building outline still cannot solve the topological errors in building outlines generated by connecting points (such as intersecting edge lines, overlapping buildings, and incomplete building outline closures). Summary of the Invention

[0004] To address the aforementioned problems, this invention proposes a method for extracting vector building outlines from remote sensing images, comprising the following steps:

[0005] S1. Acquire remote sensing images of buildings and divide the remote sensing images into training datasets and test datasets. The training dataset includes remote sensing images and building corner coordinate information, and the test dataset includes remote sensing images.

[0006] S2. Construct a building vector contour extraction model, which includes a feature pyramid network, a region proposal network, a Transformer network, a recurrent neural network, a Fourier transform module, and a feature fusion module;

[0007] S3. Train the model using the training dataset to obtain the building's outline information. Optimize the loss function in the outline information by iterating through the model's parameters until the building vector outline extraction model converges, resulting in the trained building vector outline extraction model. Specifically:

[0008] S31. Input the training dataset into the feature pyramid network to obtain the multi-dimensional feature map of the entire remote sensing image;

[0009] S32. Input the multi-dimensional feature map into the region proposal network to generate the location coordinates of buildings in the remote sensing image;

[0010] S33. Input the location coordinates of a single building into the Transformer network to generate preliminary outline information of the building, including: detailed overall outline features of the building, key corner features of the building, and edge features of the building;

[0011] S34. Take the maximum value point in the key corner feature as the starting corner point of the building edge to obtain the point coordinates. Input the point coordinates, the key corner feature of the building, and the building edge feature into the recurrent neural network to obtain the position feature of the next node.

[0012] S35. Input the overall outline features of the building into the Fourier transform module. The overall outline features of the building are converted into boundary point coordinates and then subjected to Fourier transform to obtain Fourier shape descriptor features.

[0013] S36. Input the position features of the next node obtained in step S34 and the Fourier shape descriptor features obtained in step S35 into the feature fusion module to generate the coordinates of the next node. Through continuous iteration, the corner points of the building edge are generated one by one to form complete building outline information.

[0014] S4. Input the test dataset into the trained building vector contour extraction model to obtain the building vector contour information.

[0015] Furthermore, the overall outline features of the building are constrained by an overall outline optimization loss function, which is:

[0016]

[0017] Where, N mask To output the number of categories, y i The image segmentation mask prediction result for the i-th class generated by the Transformer network. The true result of the cut mask for class i.

[0018] Furthermore, the key corner features of the building are constrained by a key corner optimization loss function, which is:

[0019]

[0020] Where, N footprints To output the number of categories, y i The generated key corner prediction results for the Transformer network. Provides accurate information for key corner points.

[0021] Furthermore, the building edge features are constrained by an edge optimization loss function, which is:

[0022]

[0023] Where, N boundart To output the number of categories, y i The edge prediction results generated by the Transformer network. This is real information at the edge.

[0024] Furthermore, step S36 generates complete building outline information, which is then constrained by a vector outline optimization loss function. The vector outline optimization loss function is as follows:

[0025]

[0026]

[0027] Among them, Y x,y For building outline labels, The Transformer network predicts the building outline, where (x, y) are the coordinates within the current building outline, Guass() is a Gaussian function used to assign specific weights to areas inside the building that are not part of the boundary, k is a hyperparameter, and A is a matrix with a value of 1, corresponding to the feature size, which sets non-boundary pixels to 1, (Y, S). + The positive portion inside the building outline refers to the set of coordinates (Y,S) of the part defined as the "building" within the building outline. - The negative part within the building outline refers to the set of coordinates of the "non-building" part within the building outline.

[0028] Furthermore, the point coordinates in step S34 change sequentially with the number of iterations as follows: (starting corner point, starting corner point, starting corner point), (starting corner point, starting corner point, next point of starting corner point), (starting corner point, next point of starting corner point, next two points of starting corner point), (starting corner point, next two points of starting corner point, next three points of starting corner point).

[0029] Furthermore, the calculation method for the Fourier shape descriptor features is as follows:

[0030]

[0031] Where D is the normalized Fourier shape descriptor feature vector, F1 is the first frequency component of the Fourier descriptor, and F... n Let n be the nth frequency component of the Fourier descriptor, and let ||| denote the norm.

[0032] This invention also proposes a remote sensing image vector building outline extraction system, comprising:

[0033] The data acquisition module is used to acquire remote sensing images of buildings and divide the remote sensing images into a training dataset and a test dataset. The training dataset includes remote sensing images and building corner coordinate information, and the test dataset includes remote sensing images.

[0034] The model building module is used to build a building vector contour extraction model, which includes a feature pyramid network, a region proposal network, a Transformer network, a recurrent neural network, a Fourier transform module, and a feature fusion module.

[0035] The model training module is used to train the model using a training dataset to obtain the building's outline information. With the goal of optimizing the loss function in the outline information, the module iterates through the model's parameters until the building vector outline extraction model converges, resulting in the trained building vector outline extraction model. Specifically:

[0036] The training dataset is input into the feature pyramid network to obtain a multi-dimensional feature map of the entire remote sensing image;

[0037] The multi-dimensional feature map is input into the region proposal network to generate the location coordinates of buildings in the remote sensing image;

[0038] The location coordinates of a single building are input into the Transformer network to generate preliminary outline information of the building, including: detailed overall outline features of the building, key corner features of the building, and edge features of the building;

[0039] The maximum value point in the key corner feature is used as the starting corner point of the building edge to obtain the point coordinates. The point coordinates, key corner features of the building, and building edge features are then input into a recurrent neural network to obtain the position features of the next node.

[0040] The overall outline features of the building are input into the Fourier transform module. The overall outline features of the building are converted into boundary point coordinates and then subjected to Fourier transform to obtain Fourier shape descriptor features.

[0041] The obtained position features of the next node and the obtained Fourier shape descriptor features are input into the feature fusion module to generate the coordinates of the next node. Through continuous iteration, the corner points of the building edge are generated one by one to form complete building outline information.

[0042] The detection module is used to input the test dataset into the trained building vector contour extraction model to obtain the building vector contour information.

[0043] The present invention also proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the remote sensing image vector contour extraction method described above.

[0044] The present invention also proposes an electronic device, including a processor and a memory, wherein the processor and the memory are interconnected, the memory is used to store a computer program, the computer program includes computer-readable instructions, and the processor is configured to invoke the computer-readable instructions to execute the above-described remote sensing image vector contour extraction method.

[0045] The beneficial effects of the technical solution provided by this invention are:

[0046] This invention constructs a building vector contour extraction model comprising a feature pyramid network, a region proposal network, a Transformer network, a recurrent neural network, a Fourier transform module, and a feature fusion module. The model is trained using a training dataset of remote sensing imagery containing building corner coordinates. The key corner features, edge features, and coordinates of the points with the maximum values ​​among the key corner features generated by the Transformer network are input into the recurrent neural network to obtain the positional features of the next node. The overall building contour features generated by the Transformer network are input into the Fourier transform module to obtain Fourier shape descriptor features. The positional features and Fourier shape descriptor features of the next node are input into the feature fusion module to generate the coordinates of the next node. Through continuous iteration, building edge corners are generated one by one, forming complete building contour information. Utilizing the building contour information and the powerful learning ability of the Transformer module for building shape features, the model improves the misjudgment of building ridge lines in remote sensing imagery, effectively constrains incomplete building shapes, and generates corresponding shape description features by obtaining Fourier descriptors from the edge points of the overall contour features. This feature effectively enhances the model's ability to perceive building outlines and accurately controls the position and orientation during the point-by-point generation process, thereby reducing topological errors in building outlines. This invention also designs a vector outline optimization loss function that constrains building shape. This loss function treats building edge corner points as samples that are difficult to classify accurately, and uses a dilated morphological method to judge all difficult-to-classify corner points, thereby reducing errors in the building edge point extraction process. Attached Figure Description

[0047] Figure 1 This is a flowchart of the method for extracting vector building outlines from remote sensing images according to an embodiment of the present invention;

[0048] Figure 2 This is a flowchart illustrating the technical process of training a building vector contour extraction model in an embodiment of the present invention.

[0049] Figure 3 This is a block diagram of an electronic device according to an exemplary embodiment of Embodiment 1 of the present invention;

[0050] Figure 4 This is a rendering of the method for extracting building outlines from remote sensing images according to an embodiment of the present invention. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0052] The flowchart of the remote sensing image vector building outline extraction method of this invention is as follows: Figure 1Specifically, it includes the following steps:

[0053] S1. Acquire remote sensing images of buildings and divide the images into training and testing datasets. The training dataset includes remote sensing images and building corner coordinate information, while the testing dataset includes remote sensing images.

[0054] In a further embodiment, the raw data of the building remote sensing image is first acquired. The storage format of the building outline in the large-scale remote sensing image in the raw data is converted from a "shapefile" file to a "JSON" format. JSON format is a lightweight text data format that is easy to understand and interpret. Compared to the binary shapefile format, JSON format data is easier to read and edit, and is more suitable for parsing and manipulating in various programming languages. The acquired remote sensing image and the corresponding building corner coordinate information are then segmented and downsampled to unify the remote sensing image size to 512*512 pixels.

[0055] S2. Construct a building vector contour extraction model, which includes a Feature Pyramid Network (FPN), a Region Proposal Network (RPN), a Transformer network, a Recurrent Neural Network (RNN), a Fourier Transform module, and a feature fusion module.

[0056] The feature pyramid network scales the input image to obtain images of different sizes, then generates feature maps of different sizes based on each size image, and finally makes predictions based on the multi-size feature maps.

[0057] The region proposal network outputs a series of ROIs (Regions of Interest) from the input feature map.

[0058] Recurrent neural networks are very effective for data with logical order, and can extract logical information from the data.

[0059] S3. Train the building vector contour extraction model constructed in S2 using the training dataset to obtain the building contour information. Optimize the loss function in the contour information by iterating through the model's parameters until the building vector contour extraction model converges, resulting in the trained building vector contour extraction model. The technical flowchart for training the building vector contour extraction model in this embodiment is shown below. Figure 2 Specifically:

[0060] S31. Input the training dataset into the feature pyramid network to obtain the multi-dimensional feature map of the entire remote sensing image;

[0061] S32. Input the multi-dimensional feature map into the region proposal network to generate the location coordinates of buildings in the remote sensing image;

[0062] S33, Set the location coordinates C of a single building n (n represents the nth building) Input into the Transformer network to generate preliminary outline information of the building, including: detailed overall outline features, key corner features, and edge features. The generated three types of features are constrained by corresponding overall outline optimization loss functions, key corner optimization loss functions, and edge optimization loss functions.

[0063] The overall outline features of the building are constrained by an overall outline optimization loss function, which is:

[0064]

[0065] Where, N mask To output the number of categories, y i The image segmentation mask prediction result for the i-th class generated by the Transformer network. The true result of the cut mask for class i.

[0066] The key corner features of a building are constrained by a key corner optimization loss function, which is as follows:

[0067]

[0068] Where, N footprints To output the number of categories, y i The generated key corner prediction results for the Transformer network. Provides accurate information for key corner points.

[0069] Building edge features are constrained by an edge optimization loss function, which is:

[0070]

[0071] Where, N boundary To output the number of categories, y i The edge prediction results generated by the Transformer network. This is real information at the edge.

[0072] S34. Take the maximum value point in the key corner feature as the starting corner point of the building edge to obtain the point coordinates. Input the point coordinates, key corner feature of the building, and building edge feature into the recurrent neural network to obtain the position feature of the next node.

[0073] S35. Input the overall outline features of the building into the Fourier transform module. The overall outline features of the building are converted into boundary point coordinates and then subjected to Fourier transform to obtain Fourier shape descriptor features.

[0074] The method for calculating the Fourier shape descriptor features is as follows:

[0075]

[0076] Among them, F1 is the first frequency component of the Fourier descriptor, which captures the basic features of the building's outline shape and represents the overall information of the entire shape; F n Let F be the nth frequency component of the Fourier descriptor. It contains more subtle local features and details of the building outline, capturing more refined shape features than F1. D represents a normalized Fourier shape descriptor feature vector, which is based on each F1. n The ratio of the modulus of F1 to the modulus of F1 is used to reveal the relative importance of different frequency components to the overall shape characteristics, and ||| represents the norm.

[0077] S36. Input the position features of the next node obtained in step S34 and the Fourier shape descriptor features obtained in step S35 into the feature fusion module, use convolution operation to generate the coordinates of the next node, and generate the building edge corner points one by one through continuous iteration to form complete building outline information.

[0078] Complete building outline information is constrained by a vector outline optimization loss function, which is:

[0079]

[0080]

[0081] Among them, Y x,y For building outline labels, This is the building outline predicted by the Transformer network, where (x, y) are the coordinates within the current building outline box. `Guass()` is a 5x5 Gaussian function used to assign specific weights to non-boundary regions within the building. `k` is a hyperparameter, and `A` is a matrix with a value of 1, corresponding to the feature size. It sets non-boundary pixels to 1. (Y, S) +The positive part inside the building outline refers to the set of coordinates of the part defined as "building" within the building outline, while (Y,S) represents the negative part inside the building outline, which refers to the set of coordinates of the part defined as "non-building" within the building outline. In this formula, they are used to calculate the input of the Gaussian function to adjust the weights of pixels or coordinates within the boundary so that the model responds to areas outside the boundary.

[0082] The point coordinates in step S34 change sequentially with the number of iterations as follows: (starting corner point, starting corner point, starting corner point), (starting corner point, starting corner point, next point of starting corner point), (starting corner point, next point of starting corner point, next two points of starting corner point), (starting corner point, next two points of starting corner point, next three points of starting corner point).

[0083] S4. Input the test dataset into the trained building vector contour extraction model to obtain the building vector contour information.

[0084] This invention also proposes a remote sensing image vector building outline extraction system, comprising:

[0085] The data acquisition module is used to acquire remote sensing images of buildings and divide the remote sensing images into a training dataset and a test dataset. The training dataset includes remote sensing images and building corner coordinate information, and the test dataset includes remote sensing images.

[0086] The model building module is used to build a building vector contour extraction model, which includes a feature pyramid network, a region proposal network, a Transformer network, a recurrent neural network, a Fourier transform module, and a feature fusion module.

[0087] The model training module is used to train the model using a training dataset to obtain the building's outline information. With the goal of optimizing the loss function in the outline information, the module iterates through the model's parameters until the building vector outline extraction model converges, resulting in the trained building vector outline extraction model. Specifically:

[0088] The training dataset is input into the feature pyramid network to obtain a multi-dimensional feature map of the entire remote sensing image;

[0089] The multi-dimensional feature map is input into the region proposal network to generate the location coordinates of buildings in the remote sensing image;

[0090] The location coordinates of a single building are input into the Transformer network to generate preliminary outline information of the building, including: detailed overall outline features of the building, key corner features of the building, and edge features of the building;

[0091] The maximum value point in the key corner feature is used as the starting corner point of the building edge to obtain the point coordinates. The point coordinates, key corner features of the building, and building edge features are then input into a recurrent neural network to obtain the position features of the next node.

[0092] The overall outline features of the building are input into the Fourier transform module. The overall outline features of the building are converted into boundary point coordinates and then subjected to Fourier transform to obtain Fourier shape descriptor features.

[0093] The obtained position features of the next node and the obtained Fourier shape descriptor features are input into the feature fusion module to generate the coordinates of the next node. Through continuous iteration, the corner points of the building edge are generated one by one to form complete building outline information.

[0094] The detection module is used to input the test dataset into the trained building vector contour extraction model to obtain the building vector contour information.

[0095] In one exemplary embodiment, a computer-readable storage medium is included, which stores a computer program that, when executed by a processor, implements the steps of the remote sensing image vector building outline extraction method described above.

[0096] Please see Figure 3 In one exemplary embodiment, the device further includes an electronic device including at least one processor, at least one memory, and at least one communication bus.

[0097] The memory stores a computer program, which includes computer-readable instructions. The processor calls the computer-readable instructions stored in the memory through the communication bus to execute the aforementioned method for extracting vector building outlines from remote sensing images.

[0098] The effect diagram of the remote sensing image vector building outline extraction method of this invention is shown in the figure below. Figure 4 .

[0099] Original image portion: In Figure 4 The upper part displays a typical original remote sensing image, which covers buildings and the natural environment of a specific area. The original image portion is intended to show remote sensing images to which embodiments of the present invention are applied.

[0100] Building outline extraction section: In Figure 4 The lower part of the image shows the result generated by the remote sensing image vector building outline extraction method provided in this embodiment of the invention. The extracted building outline clearly depicts the outline boundaries of each building in the original image. The extracted outlines have high precision and accuracy, reflecting the superior performance of this method under high resolution remote sensing image conditions.

[0101] By comparing the original image and the extracted building outline image, the superiority of the present invention is clearly evident. The extracted building outline accurately depicts the shape of the building, providing more refined and flexible geographic information. This has significant practical application value for urban planning, Geographic Information System (GIS) applications, and other related fields.

[0102] The above detailed description and renderings clearly demonstrate the innovation and practicality of the present invention in building outline extraction.

[0103] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system 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 system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0104] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for extracting vector building outlines from remote sensing images, characterized in that, Includes the following steps: S1. Acquire remote sensing images of buildings and divide the remote sensing images into training datasets and test datasets. The training dataset includes remote sensing images and building corner coordinate information, and the test dataset includes remote sensing images. S2. Construct a building vector contour extraction model, which includes a feature pyramid network, a region proposal network, a Transformer network, a recurrent neural network, a Fourier transform module, and a feature fusion module; S3. Train the model using the training dataset to obtain the building's outline information. Optimize the loss function in the outline information by iterating through the model's parameters until the building vector outline extraction model converges, resulting in the trained building vector outline extraction model. Specifically: S31. Input the training dataset into the feature pyramid network to obtain the multi-dimensional feature map of the entire remote sensing image; S32. Input the multi-dimensional feature map into the region proposal network to generate the location coordinates of buildings in the remote sensing image; S33. Input the location coordinates of a single building into the Transformer network to generate preliminary outline information of the building, including: detailed overall outline features of the building, key corner features of the building, and edge features of the building; S34. Take the maximum value point in the key corner feature as the starting corner point of the building edge to obtain the point coordinates. Input the point coordinates, the key corner feature of the building, and the building edge feature into the recurrent neural network to obtain the position feature of the next node. S35. Input the overall outline features of the building into the Fourier transform module. The overall outline features of the building are transformed into boundary point coordinates and then Fourier transform is performed to obtain the Fourier shape descriptor features. S36. Input the position features of the next node obtained in step S34 and the Fourier shape descriptor features obtained in step S35 into the feature fusion module to generate the coordinates of the next node. Through continuous iteration, the corner points of the building edge are generated one by one to form complete building outline information. The formation of complete building outline information is constrained by the vector outline optimization loss function. S4. Input the test dataset into the trained building vector contour extraction model to obtain the building vector contour information.

2. The method for extracting vector building outlines from remote sensing images according to claim 1, characterized in that, The overall outline features of the building are constrained by an overall outline optimization loss function, which is: in, To output the number of categories, The image segmentation mask prediction result for the i-th class generated by the Transformer network. This represents the actual result of the segmentation mask for the i-th class.

3. The method for extracting vector building outlines from remote sensing images according to claim 1, characterized in that, The key corner features of a building are constrained by a key corner optimization loss function, which is as follows: in, To output the number of categories, The key corner prediction results generated by the Transformer network. Provides accurate information for key corner points.

4. The method for extracting vector building outlines from remote sensing images according to claim 1, characterized in that, Building edge features are constrained by an edge optimization loss function, which is: in, To output the number of categories, Edge prediction results generated by the Transformer network This is real information at the edge.

5. The method for extracting vector building outlines from remote sensing images according to claim 1, characterized in that, The loss function for vector contour optimization is: in, For building outline labels, The building outline predicted by the Transformer network. The coordinates are within the current building outline. `Guass()` is a Gaussian function used to assign specific weight values ​​to areas inside the building that are not part of the boundary. This is a hyperparameter, where A is a matrix with a value of 1, corresponding to the feature size. It sets non-boundary pixels to 1. The positive portion inside the building's outline refers to the set of coordinates of the part defined as the "building" within the building's outline. The negative part of the building outline refers to the set of coordinates of the part of the building outline that is defined as "non-building".

6. The method for extracting vector building outlines from remote sensing images according to claim 1, characterized in that, The point coordinates in step S34 change sequentially with the number of iterations as follows: (starting corner point, starting corner point, starting corner point), (starting corner point, starting corner point, next point of starting corner point), (starting corner point, next point of starting corner point, next two points of starting corner point), (starting corner point, next two points of starting corner point, next three points of starting corner point).

7. The method for extracting vector building outlines from remote sensing images according to claim 1, characterized in that, The method for calculating the Fourier shape descriptor features is as follows: Where D is the normalized Fourier shape descriptor feature vector. This is the first frequency component of the Fourier descriptor. Let n be the nth frequency component of the Fourier descriptor. Represents the norm.

8. A system for extracting vector building outlines from remote sensing images, characterized in that, include: The data acquisition module is used to acquire remote sensing images of buildings and divide the remote sensing images into a training dataset and a test dataset. The training dataset includes remote sensing images and building corner coordinate information, and the test dataset includes remote sensing images. The model building module is used to build a building vector contour extraction model, which includes a feature pyramid network, a region proposal network, a Transformer network, a recurrent neural network, a Fourier transform module, and a feature fusion module. The model training module is used to train the model using a training dataset to obtain the building's outline information. With the goal of optimizing the loss function in the outline information, the module iterates through the model's parameters until the building vector outline extraction model converges, resulting in the trained building vector outline extraction model. Specifically: The training dataset is input into the feature pyramid network to obtain a multi-dimensional feature map of the entire remote sensing image; The multi-dimensional feature map is input into the region proposal network to generate the location coordinates of buildings in the remote sensing image; The location coordinates of a single building are input into the Transformer network to generate preliminary outline information of the building, including: detailed overall outline features of the building, key corner features of the building, and edge features of the building; The maximum value point in the key corner feature is used as the starting corner point of the building edge to obtain the point coordinates. The point coordinates, key corner features of the building, and building edge features are then input into a recurrent neural network to obtain the position features of the next node. The overall outline features of the building are input into the Fourier transform module, where they are converted into boundary point coordinates and then subjected to Fourier transform to obtain Fourier shape descriptor features. The obtained position features of the next node and the obtained Fourier shape descriptor features are input into the feature fusion module to generate the coordinates of the next node. Through continuous iteration, the corner points of the building edge are generated one by one to form complete building outline information. The formation of complete building outline information is constrained by the vector outline optimization loss function. The detection module is used to input the test dataset into the trained building vector contour extraction model to obtain the building vector contour information.

9. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1-7.

10. An electronic device, characterized in that, The device includes a processor and a memory interconnected thereto, wherein the memory is used to store a computer program, the computer program including computer-readable instructions, and the processor is configured to invoke the computer-readable instructions to perform the method as described in any one of claims 1-7.

Images